linux_dsm_epyc7002/fs/f2fs/segment.c

4013 lines
101 KiB
C
Raw Normal View History

/*
* fs/f2fs/segment.c
*
* Copyright (c) 2012 Samsung Electronics Co., Ltd.
* http://www.samsung.com/
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License version 2 as
* published by the Free Software Foundation.
*/
#include <linux/fs.h>
#include <linux/f2fs_fs.h>
#include <linux/bio.h>
#include <linux/blkdev.h>
#include <linux/prefetch.h>
#include <linux/kthread.h>
#include <linux/swap.h>
#include <linux/timer.h>
#include <linux/freezer.h>
#include <linux/sched/signal.h>
#include "f2fs.h"
#include "segment.h"
#include "node.h"
#include "gc.h"
#include "trace.h"
#include <trace/events/f2fs.h>
#define __reverse_ffz(x) __reverse_ffs(~(x))
static struct kmem_cache *discard_entry_slab;
static struct kmem_cache *discard_cmd_slab;
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
static struct kmem_cache *sit_entry_set_slab;
static struct kmem_cache *inmem_entry_slab;
static unsigned long __reverse_ulong(unsigned char *str)
{
unsigned long tmp = 0;
int shift = 24, idx = 0;
#if BITS_PER_LONG == 64
shift = 56;
#endif
while (shift >= 0) {
tmp |= (unsigned long)str[idx++] << shift;
shift -= BITS_PER_BYTE;
}
return tmp;
}
/*
* __reverse_ffs is copied from include/asm-generic/bitops/__ffs.h since
* MSB and LSB are reversed in a byte by f2fs_set_bit.
*/
static inline unsigned long __reverse_ffs(unsigned long word)
{
int num = 0;
#if BITS_PER_LONG == 64
if ((word & 0xffffffff00000000UL) == 0)
num += 32;
else
word >>= 32;
#endif
if ((word & 0xffff0000) == 0)
num += 16;
else
word >>= 16;
if ((word & 0xff00) == 0)
num += 8;
else
word >>= 8;
if ((word & 0xf0) == 0)
num += 4;
else
word >>= 4;
if ((word & 0xc) == 0)
num += 2;
else
word >>= 2;
if ((word & 0x2) == 0)
num += 1;
return num;
}
/*
* __find_rev_next(_zero)_bit is copied from lib/find_next_bit.c because
* f2fs_set_bit makes MSB and LSB reversed in a byte.
* @size must be integral times of unsigned long.
* Example:
* MSB <--> LSB
* f2fs_set_bit(0, bitmap) => 1000 0000
* f2fs_set_bit(7, bitmap) => 0000 0001
*/
static unsigned long __find_rev_next_bit(const unsigned long *addr,
unsigned long size, unsigned long offset)
{
const unsigned long *p = addr + BIT_WORD(offset);
unsigned long result = size;
unsigned long tmp;
if (offset >= size)
return size;
size -= (offset & ~(BITS_PER_LONG - 1));
offset %= BITS_PER_LONG;
while (1) {
if (*p == 0)
goto pass;
tmp = __reverse_ulong((unsigned char *)p);
tmp &= ~0UL >> offset;
if (size < BITS_PER_LONG)
tmp &= (~0UL << (BITS_PER_LONG - size));
if (tmp)
goto found;
pass:
if (size <= BITS_PER_LONG)
break;
size -= BITS_PER_LONG;
offset = 0;
p++;
}
return result;
found:
return result - size + __reverse_ffs(tmp);
}
static unsigned long __find_rev_next_zero_bit(const unsigned long *addr,
unsigned long size, unsigned long offset)
{
const unsigned long *p = addr + BIT_WORD(offset);
unsigned long result = size;
unsigned long tmp;
if (offset >= size)
return size;
size -= (offset & ~(BITS_PER_LONG - 1));
offset %= BITS_PER_LONG;
while (1) {
if (*p == ~0UL)
goto pass;
tmp = __reverse_ulong((unsigned char *)p);
if (offset)
tmp |= ~0UL << (BITS_PER_LONG - offset);
if (size < BITS_PER_LONG)
tmp |= ~0UL >> size;
if (tmp != ~0UL)
goto found;
pass:
if (size <= BITS_PER_LONG)
break;
size -= BITS_PER_LONG;
offset = 0;
p++;
}
return result;
found:
return result - size + __reverse_ffz(tmp);
}
bool need_SSR(struct f2fs_sb_info *sbi)
{
int node_secs = get_blocktype_secs(sbi, F2FS_DIRTY_NODES);
int dent_secs = get_blocktype_secs(sbi, F2FS_DIRTY_DENTS);
int imeta_secs = get_blocktype_secs(sbi, F2FS_DIRTY_IMETA);
if (test_opt(sbi, LFS))
return false;
if (sbi->gc_thread && sbi->gc_thread->gc_urgent)
return true;
return free_sections(sbi) <= (node_secs + 2 * dent_secs + imeta_secs +
SM_I(sbi)->min_ssr_sections + reserved_sections(sbi));
}
void register_inmem_page(struct inode *inode, struct page *page)
{
struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
struct f2fs_inode_info *fi = F2FS_I(inode);
struct inmem_pages *new;
f2fs_trace_pid(page);
set_page_private(page, (unsigned long)ATOMIC_WRITTEN_PAGE);
SetPagePrivate(page);
new = f2fs_kmem_cache_alloc(inmem_entry_slab, GFP_NOFS);
/* add atomic page indices to the list */
new->page = page;
INIT_LIST_HEAD(&new->list);
/* increase reference count with clean state */
mutex_lock(&fi->inmem_lock);
get_page(page);
list_add_tail(&new->list, &fi->inmem_pages);
spin_lock(&sbi->inode_lock[ATOMIC_FILE]);
if (list_empty(&fi->inmem_ilist))
list_add_tail(&fi->inmem_ilist, &sbi->inode_list[ATOMIC_FILE]);
spin_unlock(&sbi->inode_lock[ATOMIC_FILE]);
inc_page_count(F2FS_I_SB(inode), F2FS_INMEM_PAGES);
mutex_unlock(&fi->inmem_lock);
trace_f2fs_register_inmem_page(page, INMEM);
}
static int __revoke_inmem_pages(struct inode *inode,
struct list_head *head, bool drop, bool recover)
{
struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
struct inmem_pages *cur, *tmp;
int err = 0;
list_for_each_entry_safe(cur, tmp, head, list) {
struct page *page = cur->page;
if (drop)
trace_f2fs_commit_inmem_page(page, INMEM_DROP);
lock_page(page);
if (recover) {
struct dnode_of_data dn;
struct node_info ni;
trace_f2fs_commit_inmem_page(page, INMEM_REVOKE);
retry:
set_new_dnode(&dn, inode, NULL, NULL, 0);
err = get_dnode_of_data(&dn, page->index, LOOKUP_NODE);
if (err) {
if (err == -ENOMEM) {
congestion_wait(BLK_RW_ASYNC, HZ/50);
cond_resched();
goto retry;
}
err = -EAGAIN;
goto next;
}
get_node_info(sbi, dn.nid, &ni);
if (cur->old_addr == NEW_ADDR) {
invalidate_blocks(sbi, dn.data_blkaddr);
f2fs_update_data_blkaddr(&dn, NEW_ADDR);
} else
f2fs_replace_block(sbi, &dn, dn.data_blkaddr,
cur->old_addr, ni.version, true, true);
f2fs_put_dnode(&dn);
}
next:
/* we don't need to invalidate this in the sccessful status */
if (drop || recover)
ClearPageUptodate(page);
set_page_private(page, 0);
ClearPagePrivate(page);
f2fs_put_page(page, 1);
list_del(&cur->list);
kmem_cache_free(inmem_entry_slab, cur);
dec_page_count(F2FS_I_SB(inode), F2FS_INMEM_PAGES);
}
return err;
}
void drop_inmem_pages_all(struct f2fs_sb_info *sbi)
{
struct list_head *head = &sbi->inode_list[ATOMIC_FILE];
struct inode *inode;
struct f2fs_inode_info *fi;
next:
spin_lock(&sbi->inode_lock[ATOMIC_FILE]);
if (list_empty(head)) {
spin_unlock(&sbi->inode_lock[ATOMIC_FILE]);
return;
}
fi = list_first_entry(head, struct f2fs_inode_info, inmem_ilist);
inode = igrab(&fi->vfs_inode);
spin_unlock(&sbi->inode_lock[ATOMIC_FILE]);
if (inode) {
drop_inmem_pages(inode);
iput(inode);
}
congestion_wait(BLK_RW_ASYNC, HZ/50);
cond_resched();
goto next;
}
void drop_inmem_pages(struct inode *inode)
{
struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
struct f2fs_inode_info *fi = F2FS_I(inode);
mutex_lock(&fi->inmem_lock);
__revoke_inmem_pages(inode, &fi->inmem_pages, true, false);
spin_lock(&sbi->inode_lock[ATOMIC_FILE]);
if (!list_empty(&fi->inmem_ilist))
list_del_init(&fi->inmem_ilist);
spin_unlock(&sbi->inode_lock[ATOMIC_FILE]);
mutex_unlock(&fi->inmem_lock);
clear_inode_flag(inode, FI_ATOMIC_FILE);
clear_inode_flag(inode, FI_HOT_DATA);
stat_dec_atomic_write(inode);
}
f2fs: fix stale ATOMIC_WRITTEN_PAGE private pointer When I forced to enable atomic operations intentionally, I could hit the below panic, since we didn't clear page->private in f2fs_invalidate_page called by file truncation. The panic occurs due to NULL mapping having page->private. BUG: unable to handle kernel paging request at ffffffffffffffff IP: drop_buffers+0x38/0xe0 PGD 5d00c067 PUD 5d00e067 PMD 0 CPU: 3 PID: 1648 Comm: fsstress Tainted: G D OE 4.10.0+ #5 Hardware name: innotek GmbH VirtualBox/VirtualBox, BIOS VirtualBox 12/01/2006 task: ffff9151952863c0 task.stack: ffffaaec40db4000 RIP: 0010:drop_buffers+0x38/0xe0 RSP: 0018:ffffaaec40db74c8 EFLAGS: 00010292 Call Trace: ? page_referenced+0x8b/0x170 try_to_free_buffers+0xc5/0xe0 try_to_release_page+0x49/0x50 shrink_page_list+0x8bc/0x9f0 shrink_inactive_list+0x1dd/0x500 ? shrink_active_list+0x2c0/0x430 shrink_node_memcg+0x5eb/0x7c0 shrink_node+0xe1/0x320 do_try_to_free_pages+0xef/0x2e0 try_to_free_pages+0xe9/0x190 __alloc_pages_slowpath+0x390/0xe70 __alloc_pages_nodemask+0x291/0x2b0 alloc_pages_current+0x95/0x140 __page_cache_alloc+0xc4/0xe0 pagecache_get_page+0xab/0x2a0 grab_cache_page_write_begin+0x20/0x40 get_read_data_page+0x2e6/0x4c0 [f2fs] ? f2fs_mark_inode_dirty_sync+0x16/0x30 [f2fs] ? truncate_data_blocks_range+0x238/0x2b0 [f2fs] get_lock_data_page+0x30/0x190 [f2fs] __exchange_data_block+0xaaf/0xf40 [f2fs] f2fs_fallocate+0x418/0xd00 [f2fs] vfs_fallocate+0x157/0x220 SyS_fallocate+0x48/0x80 Signed-off-by: Yunlei He <heyunlei@huawei.com> Signed-off-by: Chao Yu <yuchao0@huawei.com> [Chao Yu: use INMEM_INVALIDATE for better tracing] Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2017-03-17 08:55:52 +07:00
void drop_inmem_page(struct inode *inode, struct page *page)
{
struct f2fs_inode_info *fi = F2FS_I(inode);
struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
struct list_head *head = &fi->inmem_pages;
struct inmem_pages *cur = NULL;
f2fs_bug_on(sbi, !IS_ATOMIC_WRITTEN_PAGE(page));
mutex_lock(&fi->inmem_lock);
list_for_each_entry(cur, head, list) {
if (cur->page == page)
break;
}
f2fs_bug_on(sbi, !cur || cur->page != page);
list_del(&cur->list);
mutex_unlock(&fi->inmem_lock);
dec_page_count(sbi, F2FS_INMEM_PAGES);
kmem_cache_free(inmem_entry_slab, cur);
ClearPageUptodate(page);
set_page_private(page, 0);
ClearPagePrivate(page);
f2fs_put_page(page, 0);
trace_f2fs_commit_inmem_page(page, INMEM_INVALIDATE);
}
static int __commit_inmem_pages(struct inode *inode,
struct list_head *revoke_list)
{
struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
struct f2fs_inode_info *fi = F2FS_I(inode);
struct inmem_pages *cur, *tmp;
struct f2fs_io_info fio = {
.sbi = sbi,
.ino = inode->i_ino,
.type = DATA,
.op = REQ_OP_WRITE,
.op_flags = REQ_SYNC | REQ_PRIO,
.io_type = FS_DATA_IO,
};
pgoff_t last_idx = ULONG_MAX;
int err = 0;
list_for_each_entry_safe(cur, tmp, &fi->inmem_pages, list) {
struct page *page = cur->page;
lock_page(page);
if (page->mapping == inode->i_mapping) {
trace_f2fs_commit_inmem_page(page, INMEM);
set_page_dirty(page);
f2fs_wait_on_page_writeback(page, DATA, true);
if (clear_page_dirty_for_io(page)) {
inode_dec_dirty_pages(inode);
remove_dirty_inode(inode);
}
retry:
fio.page = page;
fio.old_blkaddr = NULL_ADDR;
fio.encrypted_page = NULL;
fio.need_lock = LOCK_DONE;
err = do_write_data_page(&fio);
if (err) {
if (err == -ENOMEM) {
congestion_wait(BLK_RW_ASYNC, HZ/50);
cond_resched();
goto retry;
}
unlock_page(page);
break;
}
/* record old blkaddr for revoking */
cur->old_addr = fio.old_blkaddr;
last_idx = page->index;
}
unlock_page(page);
list_move_tail(&cur->list, revoke_list);
}
if (last_idx != ULONG_MAX)
f2fs_submit_merged_write_cond(sbi, inode, 0, last_idx, DATA);
if (!err)
__revoke_inmem_pages(inode, revoke_list, false, false);
return err;
}
int commit_inmem_pages(struct inode *inode)
{
struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
struct f2fs_inode_info *fi = F2FS_I(inode);
struct list_head revoke_list;
int err;
INIT_LIST_HEAD(&revoke_list);
f2fs_balance_fs(sbi, true);
f2fs_lock_op(sbi);
set_inode_flag(inode, FI_ATOMIC_COMMIT);
mutex_lock(&fi->inmem_lock);
err = __commit_inmem_pages(inode, &revoke_list);
if (err) {
int ret;
/*
* try to revoke all committed pages, but still we could fail
* due to no memory or other reason, if that happened, EAGAIN
* will be returned, which means in such case, transaction is
* already not integrity, caller should use journal to do the
* recovery or rewrite & commit last transaction. For other
* error number, revoking was done by filesystem itself.
*/
ret = __revoke_inmem_pages(inode, &revoke_list, false, true);
if (ret)
err = ret;
/* drop all uncommitted pages */
__revoke_inmem_pages(inode, &fi->inmem_pages, true, false);
}
spin_lock(&sbi->inode_lock[ATOMIC_FILE]);
if (!list_empty(&fi->inmem_ilist))
list_del_init(&fi->inmem_ilist);
spin_unlock(&sbi->inode_lock[ATOMIC_FILE]);
mutex_unlock(&fi->inmem_lock);
clear_inode_flag(inode, FI_ATOMIC_COMMIT);
f2fs_unlock_op(sbi);
return err;
}
/*
* This function balances dirty node and dentry pages.
* In addition, it controls garbage collection.
*/
void f2fs_balance_fs(struct f2fs_sb_info *sbi, bool need)
{
#ifdef CONFIG_F2FS_FAULT_INJECTION
if (time_to_inject(sbi, FAULT_CHECKPOINT)) {
f2fs_show_injection_info(FAULT_CHECKPOINT);
f2fs_stop_checkpoint(sbi, false);
}
#endif
/* balance_fs_bg is able to be pending */
if (need && excess_cached_nats(sbi))
f2fs_balance_fs_bg(sbi);
/*
* We should do GC or end up with checkpoint, if there are so many dirty
* dir/node pages without enough free segments.
*/
if (has_not_enough_free_secs(sbi, 0, 0)) {
mutex_lock(&sbi->gc_mutex);
f2fs_gc(sbi, false, false, NULL_SEGNO);
}
}
void f2fs_balance_fs_bg(struct f2fs_sb_info *sbi)
{
/* try to shrink extent cache when there is no enough memory */
if (!available_free_memory(sbi, EXTENT_CACHE))
f2fs_shrink_extent_tree(sbi, EXTENT_CACHE_SHRINK_NUMBER);
/* check the # of cached NAT entries */
if (!available_free_memory(sbi, NAT_ENTRIES))
try_to_free_nats(sbi, NAT_ENTRY_PER_BLOCK);
if (!available_free_memory(sbi, FREE_NIDS))
try_to_free_nids(sbi, MAX_FREE_NIDS);
else
build_free_nids(sbi, false, false);
if (!is_idle(sbi) && !excess_dirty_nats(sbi))
return;
/* checkpoint is the only way to shrink partial cached entries */
if (!available_free_memory(sbi, NAT_ENTRIES) ||
!available_free_memory(sbi, INO_ENTRIES) ||
excess_prefree_segs(sbi) ||
excess_dirty_nats(sbi) ||
f2fs_time_over(sbi, CP_TIME)) {
if (test_opt(sbi, DATA_FLUSH)) {
struct blk_plug plug;
blk_start_plug(&plug);
sync_dirty_inodes(sbi, FILE_INODE);
blk_finish_plug(&plug);
}
f2fs_sync_fs(sbi->sb, true);
stat_inc_bg_cp_count(sbi->stat_info);
}
}
static int __submit_flush_wait(struct f2fs_sb_info *sbi,
struct block_device *bdev)
{
struct bio *bio = f2fs_bio_alloc(sbi, 0, true);
int ret;
bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_PREFLUSH;
bio_set_dev(bio, bdev);
ret = submit_bio_wait(bio);
bio_put(bio);
trace_f2fs_issue_flush(bdev, test_opt(sbi, NOBARRIER),
test_opt(sbi, FLUSH_MERGE), ret);
return ret;
}
static int submit_flush_wait(struct f2fs_sb_info *sbi, nid_t ino)
{
int ret = 0;
int i;
if (!sbi->s_ndevs)
return __submit_flush_wait(sbi, sbi->sb->s_bdev);
for (i = 0; i < sbi->s_ndevs; i++) {
if (!is_dirty_device(sbi, ino, i, FLUSH_INO))
continue;
ret = __submit_flush_wait(sbi, FDEV(i).bdev);
if (ret)
break;
}
return ret;
}
static int issue_flush_thread(void *data)
{
struct f2fs_sb_info *sbi = data;
struct flush_cmd_control *fcc = SM_I(sbi)->fcc_info;
wait_queue_head_t *q = &fcc->flush_wait_queue;
repeat:
if (kthread_should_stop())
return 0;
f2fs: make background threads of f2fs being aware of freezing When ->freeze_fs is called from lvm for doing snapshot, it needs to make sure there will be no more changes in filesystem's data, however, previously, background threads like GC thread wasn't aware of freezing, so in environment with active background threads, data of snapshot becomes unstable. This patch fixes this issue by adding sb_{start,end}_intwrite in below background threads: - GC thread - flush thread - discard thread Note that, don't use sb_start_intwrite() in gc_thread_func() due to: generic/241 reports below bug: ====================================================== WARNING: possible circular locking dependency detected 4.13.0-rc1+ #32 Tainted: G O ------------------------------------------------------ f2fs_gc-250:0/22186 is trying to acquire lock: (&sbi->gc_mutex){+.+...}, at: [<f8fa7f0b>] f2fs_sync_fs+0x7b/0x1b0 [f2fs] but task is already holding lock: (sb_internal#2){++++.-}, at: [<f8fb5609>] gc_thread_func+0x159/0x4a0 [f2fs] which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (sb_internal#2){++++.-}: __lock_acquire+0x405/0x7b0 lock_acquire+0xae/0x220 __sb_start_write+0x11d/0x1f0 f2fs_evict_inode+0x2d6/0x4e0 [f2fs] evict+0xa8/0x170 iput+0x1fb/0x2c0 f2fs_sync_inode_meta+0x3f/0xf0 [f2fs] write_checkpoint+0x1b1/0x750 [f2fs] f2fs_sync_fs+0x85/0x1b0 [f2fs] f2fs_do_sync_file.isra.24+0x137/0xa30 [f2fs] f2fs_sync_file+0x34/0x40 [f2fs] vfs_fsync_range+0x4a/0xa0 do_fsync+0x3c/0x60 SyS_fdatasync+0x15/0x20 do_fast_syscall_32+0xa1/0x1b0 entry_SYSENTER_32+0x4c/0x7b -> #1 (&sbi->cp_mutex){+.+...}: __lock_acquire+0x405/0x7b0 lock_acquire+0xae/0x220 __mutex_lock+0x4f/0x830 mutex_lock_nested+0x25/0x30 write_checkpoint+0x2f/0x750 [f2fs] f2fs_sync_fs+0x85/0x1b0 [f2fs] sync_filesystem+0x67/0x80 generic_shutdown_super+0x27/0x100 kill_block_super+0x22/0x50 kill_f2fs_super+0x3a/0x40 [f2fs] deactivate_locked_super+0x3d/0x70 deactivate_super+0x40/0x60 cleanup_mnt+0x39/0x70 __cleanup_mnt+0x10/0x20 task_work_run+0x69/0x80 exit_to_usermode_loop+0x57/0x92 do_fast_syscall_32+0x18c/0x1b0 entry_SYSENTER_32+0x4c/0x7b -> #0 (&sbi->gc_mutex){+.+...}: validate_chain.isra.36+0xc50/0xdb0 __lock_acquire+0x405/0x7b0 lock_acquire+0xae/0x220 __mutex_lock+0x4f/0x830 mutex_lock_nested+0x25/0x30 f2fs_sync_fs+0x7b/0x1b0 [f2fs] f2fs_balance_fs_bg+0xb9/0x200 [f2fs] gc_thread_func+0x302/0x4a0 [f2fs] kthread+0xe9/0x120 ret_from_fork+0x19/0x24 other info that might help us debug this: Chain exists of: &sbi->gc_mutex --> &sbi->cp_mutex --> sb_internal#2 Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(sb_internal#2); lock(&sbi->cp_mutex); lock(sb_internal#2); lock(&sbi->gc_mutex); *** DEADLOCK *** 1 lock held by f2fs_gc-250:0/22186: #0: (sb_internal#2){++++.-}, at: [<f8fb5609>] gc_thread_func+0x159/0x4a0 [f2fs] stack backtrace: CPU: 2 PID: 22186 Comm: f2fs_gc-250:0 Tainted: G O 4.13.0-rc1+ #32 Hardware name: innotek GmbH VirtualBox/VirtualBox, BIOS VirtualBox 12/01/2006 Call Trace: dump_stack+0x5f/0x92 print_circular_bug+0x1b3/0x1bd validate_chain.isra.36+0xc50/0xdb0 ? __this_cpu_preempt_check+0xf/0x20 __lock_acquire+0x405/0x7b0 lock_acquire+0xae/0x220 ? f2fs_sync_fs+0x7b/0x1b0 [f2fs] __mutex_lock+0x4f/0x830 ? f2fs_sync_fs+0x7b/0x1b0 [f2fs] mutex_lock_nested+0x25/0x30 ? f2fs_sync_fs+0x7b/0x1b0 [f2fs] f2fs_sync_fs+0x7b/0x1b0 [f2fs] f2fs_balance_fs_bg+0xb9/0x200 [f2fs] gc_thread_func+0x302/0x4a0 [f2fs] ? preempt_schedule_common+0x2f/0x4d ? f2fs_gc+0x540/0x540 [f2fs] kthread+0xe9/0x120 ? f2fs_gc+0x540/0x540 [f2fs] ? kthread_create_on_node+0x30/0x30 ret_from_fork+0x19/0x24 The deadlock occurs in below condition: GC Thread Thread B - sb_start_intwrite - f2fs_sync_file - f2fs_sync_fs - mutex_lock(&sbi->gc_mutex) - write_checkpoint - block_operations - f2fs_sync_inode_meta - iput - sb_start_intwrite - mutex_lock(&sbi->gc_mutex) Fix this by altering sb_start_intwrite to sb_start_write_trylock. Signed-off-by: Chao Yu <yuchao0@huawei.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2017-07-22 07:52:23 +07:00
sb_start_intwrite(sbi->sb);
if (!llist_empty(&fcc->issue_list)) {
struct flush_cmd *cmd, *next;
int ret;
fcc->dispatch_list = llist_del_all(&fcc->issue_list);
fcc->dispatch_list = llist_reverse_order(fcc->dispatch_list);
cmd = llist_entry(fcc->dispatch_list, struct flush_cmd, llnode);
ret = submit_flush_wait(sbi, cmd->ino);
atomic_inc(&fcc->issued_flush);
llist_for_each_entry_safe(cmd, next,
fcc->dispatch_list, llnode) {
cmd->ret = ret;
complete(&cmd->wait);
}
fcc->dispatch_list = NULL;
}
f2fs: make background threads of f2fs being aware of freezing When ->freeze_fs is called from lvm for doing snapshot, it needs to make sure there will be no more changes in filesystem's data, however, previously, background threads like GC thread wasn't aware of freezing, so in environment with active background threads, data of snapshot becomes unstable. This patch fixes this issue by adding sb_{start,end}_intwrite in below background threads: - GC thread - flush thread - discard thread Note that, don't use sb_start_intwrite() in gc_thread_func() due to: generic/241 reports below bug: ====================================================== WARNING: possible circular locking dependency detected 4.13.0-rc1+ #32 Tainted: G O ------------------------------------------------------ f2fs_gc-250:0/22186 is trying to acquire lock: (&sbi->gc_mutex){+.+...}, at: [<f8fa7f0b>] f2fs_sync_fs+0x7b/0x1b0 [f2fs] but task is already holding lock: (sb_internal#2){++++.-}, at: [<f8fb5609>] gc_thread_func+0x159/0x4a0 [f2fs] which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (sb_internal#2){++++.-}: __lock_acquire+0x405/0x7b0 lock_acquire+0xae/0x220 __sb_start_write+0x11d/0x1f0 f2fs_evict_inode+0x2d6/0x4e0 [f2fs] evict+0xa8/0x170 iput+0x1fb/0x2c0 f2fs_sync_inode_meta+0x3f/0xf0 [f2fs] write_checkpoint+0x1b1/0x750 [f2fs] f2fs_sync_fs+0x85/0x1b0 [f2fs] f2fs_do_sync_file.isra.24+0x137/0xa30 [f2fs] f2fs_sync_file+0x34/0x40 [f2fs] vfs_fsync_range+0x4a/0xa0 do_fsync+0x3c/0x60 SyS_fdatasync+0x15/0x20 do_fast_syscall_32+0xa1/0x1b0 entry_SYSENTER_32+0x4c/0x7b -> #1 (&sbi->cp_mutex){+.+...}: __lock_acquire+0x405/0x7b0 lock_acquire+0xae/0x220 __mutex_lock+0x4f/0x830 mutex_lock_nested+0x25/0x30 write_checkpoint+0x2f/0x750 [f2fs] f2fs_sync_fs+0x85/0x1b0 [f2fs] sync_filesystem+0x67/0x80 generic_shutdown_super+0x27/0x100 kill_block_super+0x22/0x50 kill_f2fs_super+0x3a/0x40 [f2fs] deactivate_locked_super+0x3d/0x70 deactivate_super+0x40/0x60 cleanup_mnt+0x39/0x70 __cleanup_mnt+0x10/0x20 task_work_run+0x69/0x80 exit_to_usermode_loop+0x57/0x92 do_fast_syscall_32+0x18c/0x1b0 entry_SYSENTER_32+0x4c/0x7b -> #0 (&sbi->gc_mutex){+.+...}: validate_chain.isra.36+0xc50/0xdb0 __lock_acquire+0x405/0x7b0 lock_acquire+0xae/0x220 __mutex_lock+0x4f/0x830 mutex_lock_nested+0x25/0x30 f2fs_sync_fs+0x7b/0x1b0 [f2fs] f2fs_balance_fs_bg+0xb9/0x200 [f2fs] gc_thread_func+0x302/0x4a0 [f2fs] kthread+0xe9/0x120 ret_from_fork+0x19/0x24 other info that might help us debug this: Chain exists of: &sbi->gc_mutex --> &sbi->cp_mutex --> sb_internal#2 Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(sb_internal#2); lock(&sbi->cp_mutex); lock(sb_internal#2); lock(&sbi->gc_mutex); *** DEADLOCK *** 1 lock held by f2fs_gc-250:0/22186: #0: (sb_internal#2){++++.-}, at: [<f8fb5609>] gc_thread_func+0x159/0x4a0 [f2fs] stack backtrace: CPU: 2 PID: 22186 Comm: f2fs_gc-250:0 Tainted: G O 4.13.0-rc1+ #32 Hardware name: innotek GmbH VirtualBox/VirtualBox, BIOS VirtualBox 12/01/2006 Call Trace: dump_stack+0x5f/0x92 print_circular_bug+0x1b3/0x1bd validate_chain.isra.36+0xc50/0xdb0 ? __this_cpu_preempt_check+0xf/0x20 __lock_acquire+0x405/0x7b0 lock_acquire+0xae/0x220 ? f2fs_sync_fs+0x7b/0x1b0 [f2fs] __mutex_lock+0x4f/0x830 ? f2fs_sync_fs+0x7b/0x1b0 [f2fs] mutex_lock_nested+0x25/0x30 ? f2fs_sync_fs+0x7b/0x1b0 [f2fs] f2fs_sync_fs+0x7b/0x1b0 [f2fs] f2fs_balance_fs_bg+0xb9/0x200 [f2fs] gc_thread_func+0x302/0x4a0 [f2fs] ? preempt_schedule_common+0x2f/0x4d ? f2fs_gc+0x540/0x540 [f2fs] kthread+0xe9/0x120 ? f2fs_gc+0x540/0x540 [f2fs] ? kthread_create_on_node+0x30/0x30 ret_from_fork+0x19/0x24 The deadlock occurs in below condition: GC Thread Thread B - sb_start_intwrite - f2fs_sync_file - f2fs_sync_fs - mutex_lock(&sbi->gc_mutex) - write_checkpoint - block_operations - f2fs_sync_inode_meta - iput - sb_start_intwrite - mutex_lock(&sbi->gc_mutex) Fix this by altering sb_start_intwrite to sb_start_write_trylock. Signed-off-by: Chao Yu <yuchao0@huawei.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2017-07-22 07:52:23 +07:00
sb_end_intwrite(sbi->sb);
wait_event_interruptible(*q,
kthread_should_stop() || !llist_empty(&fcc->issue_list));
goto repeat;
}
int f2fs_issue_flush(struct f2fs_sb_info *sbi, nid_t ino)
{
struct flush_cmd_control *fcc = SM_I(sbi)->fcc_info;
struct flush_cmd cmd;
int ret;
if (test_opt(sbi, NOBARRIER))
return 0;
if (!test_opt(sbi, FLUSH_MERGE)) {
ret = submit_flush_wait(sbi, ino);
atomic_inc(&fcc->issued_flush);
return ret;
}
if (atomic_inc_return(&fcc->issing_flush) == 1 || sbi->s_ndevs > 1) {
ret = submit_flush_wait(sbi, ino);
atomic_dec(&fcc->issing_flush);
atomic_inc(&fcc->issued_flush);
return ret;
}
cmd.ino = ino;
init_completion(&cmd.wait);
llist_add(&cmd.llnode, &fcc->issue_list);
/* update issue_list before we wake up issue_flush thread */
smp_mb();
if (waitqueue_active(&fcc->flush_wait_queue))
wake_up(&fcc->flush_wait_queue);
if (fcc->f2fs_issue_flush) {
wait_for_completion(&cmd.wait);
atomic_dec(&fcc->issing_flush);
} else {
struct llist_node *list;
list = llist_del_all(&fcc->issue_list);
if (!list) {
wait_for_completion(&cmd.wait);
atomic_dec(&fcc->issing_flush);
} else {
struct flush_cmd *tmp, *next;
ret = submit_flush_wait(sbi, ino);
llist_for_each_entry_safe(tmp, next, list, llnode) {
if (tmp == &cmd) {
cmd.ret = ret;
atomic_dec(&fcc->issing_flush);
continue;
}
tmp->ret = ret;
complete(&tmp->wait);
}
}
}
return cmd.ret;
}
int create_flush_cmd_control(struct f2fs_sb_info *sbi)
{
dev_t dev = sbi->sb->s_bdev->bd_dev;
struct flush_cmd_control *fcc;
int err = 0;
if (SM_I(sbi)->fcc_info) {
fcc = SM_I(sbi)->fcc_info;
if (fcc->f2fs_issue_flush)
return err;
goto init_thread;
}
fcc = f2fs_kzalloc(sbi, sizeof(struct flush_cmd_control), GFP_KERNEL);
if (!fcc)
return -ENOMEM;
atomic_set(&fcc->issued_flush, 0);
atomic_set(&fcc->issing_flush, 0);
init_waitqueue_head(&fcc->flush_wait_queue);
init_llist_head(&fcc->issue_list);
SM_I(sbi)->fcc_info = fcc;
if (!test_opt(sbi, FLUSH_MERGE))
return err;
init_thread:
fcc->f2fs_issue_flush = kthread_run(issue_flush_thread, sbi,
"f2fs_flush-%u:%u", MAJOR(dev), MINOR(dev));
if (IS_ERR(fcc->f2fs_issue_flush)) {
err = PTR_ERR(fcc->f2fs_issue_flush);
kfree(fcc);
SM_I(sbi)->fcc_info = NULL;
return err;
}
return err;
}
void destroy_flush_cmd_control(struct f2fs_sb_info *sbi, bool free)
{
struct flush_cmd_control *fcc = SM_I(sbi)->fcc_info;
if (fcc && fcc->f2fs_issue_flush) {
struct task_struct *flush_thread = fcc->f2fs_issue_flush;
fcc->f2fs_issue_flush = NULL;
kthread_stop(flush_thread);
}
if (free) {
kfree(fcc);
SM_I(sbi)->fcc_info = NULL;
}
}
int f2fs_flush_device_cache(struct f2fs_sb_info *sbi)
{
int ret = 0, i;
if (!sbi->s_ndevs)
return 0;
for (i = 1; i < sbi->s_ndevs; i++) {
if (!f2fs_test_bit(i, (char *)&sbi->dirty_device))
continue;
ret = __submit_flush_wait(sbi, FDEV(i).bdev);
if (ret)
break;
spin_lock(&sbi->dev_lock);
f2fs_clear_bit(i, (char *)&sbi->dirty_device);
spin_unlock(&sbi->dev_lock);
}
return ret;
}
static void __locate_dirty_segment(struct f2fs_sb_info *sbi, unsigned int segno,
enum dirty_type dirty_type)
{
struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
/* need not be added */
if (IS_CURSEG(sbi, segno))
return;
if (!test_and_set_bit(segno, dirty_i->dirty_segmap[dirty_type]))
dirty_i->nr_dirty[dirty_type]++;
if (dirty_type == DIRTY) {
struct seg_entry *sentry = get_seg_entry(sbi, segno);
enum dirty_type t = sentry->type;
if (unlikely(t >= DIRTY)) {
f2fs_bug_on(sbi, 1);
return;
}
if (!test_and_set_bit(segno, dirty_i->dirty_segmap[t]))
dirty_i->nr_dirty[t]++;
}
}
static void __remove_dirty_segment(struct f2fs_sb_info *sbi, unsigned int segno,
enum dirty_type dirty_type)
{
struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
if (test_and_clear_bit(segno, dirty_i->dirty_segmap[dirty_type]))
dirty_i->nr_dirty[dirty_type]--;
if (dirty_type == DIRTY) {
struct seg_entry *sentry = get_seg_entry(sbi, segno);
enum dirty_type t = sentry->type;
if (test_and_clear_bit(segno, dirty_i->dirty_segmap[t]))
dirty_i->nr_dirty[t]--;
if (get_valid_blocks(sbi, segno, true) == 0)
clear_bit(GET_SEC_FROM_SEG(sbi, segno),
dirty_i->victim_secmap);
}
}
/*
* Should not occur error such as -ENOMEM.
* Adding dirty entry into seglist is not critical operation.
* If a given segment is one of current working segments, it won't be added.
*/
static void locate_dirty_segment(struct f2fs_sb_info *sbi, unsigned int segno)
{
struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
unsigned short valid_blocks;
if (segno == NULL_SEGNO || IS_CURSEG(sbi, segno))
return;
mutex_lock(&dirty_i->seglist_lock);
valid_blocks = get_valid_blocks(sbi, segno, false);
if (valid_blocks == 0) {
__locate_dirty_segment(sbi, segno, PRE);
__remove_dirty_segment(sbi, segno, DIRTY);
} else if (valid_blocks < sbi->blocks_per_seg) {
__locate_dirty_segment(sbi, segno, DIRTY);
} else {
/* Recovery routine with SSR needs this */
__remove_dirty_segment(sbi, segno, DIRTY);
}
mutex_unlock(&dirty_i->seglist_lock);
}
static struct discard_cmd *__create_discard_cmd(struct f2fs_sb_info *sbi,
struct block_device *bdev, block_t lstart,
block_t start, block_t len)
{
struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
struct list_head *pend_list;
struct discard_cmd *dc;
f2fs_bug_on(sbi, !len);
pend_list = &dcc->pend_list[plist_idx(len)];
dc = f2fs_kmem_cache_alloc(discard_cmd_slab, GFP_NOFS);
INIT_LIST_HEAD(&dc->list);
dc->bdev = bdev;
dc->lstart = lstart;
dc->start = start;
dc->len = len;
dc->ref = 0;
dc->state = D_PREP;
dc->error = 0;
init_completion(&dc->wait);
list_add_tail(&dc->list, pend_list);
atomic_inc(&dcc->discard_cmd_cnt);
dcc->undiscard_blks += len;
return dc;
}
static struct discard_cmd *__attach_discard_cmd(struct f2fs_sb_info *sbi,
struct block_device *bdev, block_t lstart,
block_t start, block_t len,
struct rb_node *parent, struct rb_node **p)
{
struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
struct discard_cmd *dc;
dc = __create_discard_cmd(sbi, bdev, lstart, start, len);
rb_link_node(&dc->rb_node, parent, p);
rb_insert_color(&dc->rb_node, &dcc->root);
return dc;
}
static void __detach_discard_cmd(struct discard_cmd_control *dcc,
struct discard_cmd *dc)
{
if (dc->state == D_DONE)
atomic_dec(&dcc->issing_discard);
list_del(&dc->list);
rb_erase(&dc->rb_node, &dcc->root);
dcc->undiscard_blks -= dc->len;
kmem_cache_free(discard_cmd_slab, dc);
atomic_dec(&dcc->discard_cmd_cnt);
}
static void __remove_discard_cmd(struct f2fs_sb_info *sbi,
struct discard_cmd *dc)
{
struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
trace_f2fs_remove_discard(dc->bdev, dc->start, dc->len);
f2fs_bug_on(sbi, dc->ref);
if (dc->error == -EOPNOTSUPP)
dc->error = 0;
if (dc->error)
f2fs_msg(sbi->sb, KERN_INFO,
"Issue discard(%u, %u, %u) failed, ret: %d",
dc->lstart, dc->start, dc->len, dc->error);
__detach_discard_cmd(dcc, dc);
}
static void f2fs_submit_discard_endio(struct bio *bio)
{
struct discard_cmd *dc = (struct discard_cmd *)bio->bi_private;
dc->error = blk_status_to_errno(bio->bi_status);
dc->state = D_DONE;
complete_all(&dc->wait);
bio_put(bio);
}
static void __check_sit_bitmap(struct f2fs_sb_info *sbi,
block_t start, block_t end)
{
#ifdef CONFIG_F2FS_CHECK_FS
struct seg_entry *sentry;
unsigned int segno;
block_t blk = start;
unsigned long offset, size, max_blocks = sbi->blocks_per_seg;
unsigned long *map;
while (blk < end) {
segno = GET_SEGNO(sbi, blk);
sentry = get_seg_entry(sbi, segno);
offset = GET_BLKOFF_FROM_SEG0(sbi, blk);
if (end < START_BLOCK(sbi, segno + 1))
size = GET_BLKOFF_FROM_SEG0(sbi, end);
else
size = max_blocks;
map = (unsigned long *)(sentry->cur_valid_map);
offset = __find_rev_next_bit(map, size, offset);
f2fs_bug_on(sbi, offset != size);
blk = START_BLOCK(sbi, segno + 1);
}
#endif
}
/* this function is copied from blkdev_issue_discard from block/blk-lib.c */
static void __submit_discard_cmd(struct f2fs_sb_info *sbi,
struct discard_policy *dpolicy,
struct discard_cmd *dc)
{
struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
struct list_head *wait_list = (dpolicy->type == DPOLICY_FSTRIM) ?
&(dcc->fstrim_list) : &(dcc->wait_list);
struct bio *bio = NULL;
int flag = dpolicy->sync ? REQ_SYNC : 0;
if (dc->state != D_PREP)
return;
trace_f2fs_issue_discard(dc->bdev, dc->start, dc->len);
dc->error = __blkdev_issue_discard(dc->bdev,
SECTOR_FROM_BLOCK(dc->start),
SECTOR_FROM_BLOCK(dc->len),
GFP_NOFS, 0, &bio);
if (!dc->error) {
/* should keep before submission to avoid D_DONE right away */
dc->state = D_SUBMIT;
atomic_inc(&dcc->issued_discard);
atomic_inc(&dcc->issing_discard);
if (bio) {
bio->bi_private = dc;
bio->bi_end_io = f2fs_submit_discard_endio;
bio->bi_opf |= flag;
submit_bio(bio);
list_move_tail(&dc->list, wait_list);
__check_sit_bitmap(sbi, dc->start, dc->start + dc->len);
f2fs_update_iostat(sbi, FS_DISCARD, 1);
}
} else {
__remove_discard_cmd(sbi, dc);
}
}
static struct discard_cmd *__insert_discard_tree(struct f2fs_sb_info *sbi,
struct block_device *bdev, block_t lstart,
block_t start, block_t len,
struct rb_node **insert_p,
struct rb_node *insert_parent)
{
struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
struct rb_node **p;
struct rb_node *parent = NULL;
struct discard_cmd *dc = NULL;
if (insert_p && insert_parent) {
parent = insert_parent;
p = insert_p;
goto do_insert;
}
p = __lookup_rb_tree_for_insert(sbi, &dcc->root, &parent, lstart);
do_insert:
dc = __attach_discard_cmd(sbi, bdev, lstart, start, len, parent, p);
if (!dc)
return NULL;
return dc;
}
static void __relocate_discard_cmd(struct discard_cmd_control *dcc,
struct discard_cmd *dc)
{
list_move_tail(&dc->list, &dcc->pend_list[plist_idx(dc->len)]);
}
static void __punch_discard_cmd(struct f2fs_sb_info *sbi,
struct discard_cmd *dc, block_t blkaddr)
{
struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
struct discard_info di = dc->di;
bool modified = false;
if (dc->state == D_DONE || dc->len == 1) {
__remove_discard_cmd(sbi, dc);
return;
}
dcc->undiscard_blks -= di.len;
if (blkaddr > di.lstart) {
dc->len = blkaddr - dc->lstart;
dcc->undiscard_blks += dc->len;
__relocate_discard_cmd(dcc, dc);
modified = true;
}
if (blkaddr < di.lstart + di.len - 1) {
if (modified) {
__insert_discard_tree(sbi, dc->bdev, blkaddr + 1,
di.start + blkaddr + 1 - di.lstart,
di.lstart + di.len - 1 - blkaddr,
NULL, NULL);
} else {
dc->lstart++;
dc->len--;
dc->start++;
dcc->undiscard_blks += dc->len;
__relocate_discard_cmd(dcc, dc);
}
}
}
static void __update_discard_tree_range(struct f2fs_sb_info *sbi,
struct block_device *bdev, block_t lstart,
block_t start, block_t len)
{
struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
struct discard_cmd *prev_dc = NULL, *next_dc = NULL;
struct discard_cmd *dc;
struct discard_info di = {0};
struct rb_node **insert_p = NULL, *insert_parent = NULL;
block_t end = lstart + len;
mutex_lock(&dcc->cmd_lock);
dc = (struct discard_cmd *)__lookup_rb_tree_ret(&dcc->root,
NULL, lstart,
(struct rb_entry **)&prev_dc,
(struct rb_entry **)&next_dc,
&insert_p, &insert_parent, true);
if (dc)
prev_dc = dc;
if (!prev_dc) {
di.lstart = lstart;
di.len = next_dc ? next_dc->lstart - lstart : len;
di.len = min(di.len, len);
di.start = start;
}
while (1) {
struct rb_node *node;
bool merged = false;
struct discard_cmd *tdc = NULL;
if (prev_dc) {
di.lstart = prev_dc->lstart + prev_dc->len;
if (di.lstart < lstart)
di.lstart = lstart;
if (di.lstart >= end)
break;
if (!next_dc || next_dc->lstart > end)
di.len = end - di.lstart;
else
di.len = next_dc->lstart - di.lstart;
di.start = start + di.lstart - lstart;
}
if (!di.len)
goto next;
if (prev_dc && prev_dc->state == D_PREP &&
prev_dc->bdev == bdev &&
__is_discard_back_mergeable(&di, &prev_dc->di)) {
prev_dc->di.len += di.len;
dcc->undiscard_blks += di.len;
__relocate_discard_cmd(dcc, prev_dc);
di = prev_dc->di;
tdc = prev_dc;
merged = true;
}
if (next_dc && next_dc->state == D_PREP &&
next_dc->bdev == bdev &&
__is_discard_front_mergeable(&di, &next_dc->di)) {
next_dc->di.lstart = di.lstart;
next_dc->di.len += di.len;
next_dc->di.start = di.start;
dcc->undiscard_blks += di.len;
__relocate_discard_cmd(dcc, next_dc);
if (tdc)
__remove_discard_cmd(sbi, tdc);
merged = true;
}
if (!merged) {
__insert_discard_tree(sbi, bdev, di.lstart, di.start,
di.len, NULL, NULL);
}
next:
prev_dc = next_dc;
if (!prev_dc)
break;
node = rb_next(&prev_dc->rb_node);
next_dc = rb_entry_safe(node, struct discard_cmd, rb_node);
}
mutex_unlock(&dcc->cmd_lock);
}
static int __queue_discard_cmd(struct f2fs_sb_info *sbi,
struct block_device *bdev, block_t blkstart, block_t blklen)
{
block_t lblkstart = blkstart;
trace_f2fs_queue_discard(bdev, blkstart, blklen);
if (sbi->s_ndevs) {
int devi = f2fs_target_device_index(sbi, blkstart);
blkstart -= FDEV(devi).start_blk;
}
__update_discard_tree_range(sbi, bdev, lblkstart, blkstart, blklen);
return 0;
}
static void __issue_discard_cmd_range(struct f2fs_sb_info *sbi,
struct discard_policy *dpolicy,
unsigned int start, unsigned int end)
{
struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
struct discard_cmd *prev_dc = NULL, *next_dc = NULL;
struct rb_node **insert_p = NULL, *insert_parent = NULL;
struct discard_cmd *dc;
struct blk_plug plug;
int issued;
next:
issued = 0;
mutex_lock(&dcc->cmd_lock);
f2fs_bug_on(sbi, !__check_rb_tree_consistence(sbi, &dcc->root));
dc = (struct discard_cmd *)__lookup_rb_tree_ret(&dcc->root,
NULL, start,
(struct rb_entry **)&prev_dc,
(struct rb_entry **)&next_dc,
&insert_p, &insert_parent, true);
if (!dc)
dc = next_dc;
blk_start_plug(&plug);
while (dc && dc->lstart <= end) {
struct rb_node *node;
if (dc->len < dpolicy->granularity)
goto skip;
if (dc->state != D_PREP) {
list_move_tail(&dc->list, &dcc->fstrim_list);
goto skip;
}
__submit_discard_cmd(sbi, dpolicy, dc);
if (++issued >= dpolicy->max_requests) {
start = dc->lstart + dc->len;
blk_finish_plug(&plug);
mutex_unlock(&dcc->cmd_lock);
schedule();
goto next;
}
skip:
node = rb_next(&dc->rb_node);
dc = rb_entry_safe(node, struct discard_cmd, rb_node);
if (fatal_signal_pending(current))
break;
}
blk_finish_plug(&plug);
mutex_unlock(&dcc->cmd_lock);
}
static int __issue_discard_cmd(struct f2fs_sb_info *sbi,
struct discard_policy *dpolicy)
{
struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
struct list_head *pend_list;
struct discard_cmd *dc, *tmp;
struct blk_plug plug;
int i, iter = 0, issued = 0;
bool io_interrupted = false;
for (i = MAX_PLIST_NUM - 1; i >= 0; i--) {
if (i + 1 < dpolicy->granularity)
break;
pend_list = &dcc->pend_list[i];
mutex_lock(&dcc->cmd_lock);
if (list_empty(pend_list))
goto next;
f2fs_bug_on(sbi, !__check_rb_tree_consistence(sbi, &dcc->root));
blk_start_plug(&plug);
list_for_each_entry_safe(dc, tmp, pend_list, list) {
f2fs_bug_on(sbi, dc->state != D_PREP);
if (dpolicy->io_aware && i < dpolicy->io_aware_gran &&
!is_idle(sbi)) {
io_interrupted = true;
goto skip;
}
__submit_discard_cmd(sbi, dpolicy, dc);
issued++;
skip:
if (++iter >= dpolicy->max_requests)
break;
}
blk_finish_plug(&plug);
next:
mutex_unlock(&dcc->cmd_lock);
if (iter >= dpolicy->max_requests)
break;
}
if (!issued && io_interrupted)
issued = -1;
return issued;
}
static bool __drop_discard_cmd(struct f2fs_sb_info *sbi)
{
struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
struct list_head *pend_list;
struct discard_cmd *dc, *tmp;
int i;
bool dropped = false;
mutex_lock(&dcc->cmd_lock);
for (i = MAX_PLIST_NUM - 1; i >= 0; i--) {
pend_list = &dcc->pend_list[i];
list_for_each_entry_safe(dc, tmp, pend_list, list) {
f2fs_bug_on(sbi, dc->state != D_PREP);
__remove_discard_cmd(sbi, dc);
dropped = true;
}
}
mutex_unlock(&dcc->cmd_lock);
return dropped;
}
void drop_discard_cmd(struct f2fs_sb_info *sbi)
{
__drop_discard_cmd(sbi);
}
static unsigned int __wait_one_discard_bio(struct f2fs_sb_info *sbi,
struct discard_cmd *dc)
{
struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
unsigned int len = 0;
wait_for_completion_io(&dc->wait);
mutex_lock(&dcc->cmd_lock);
f2fs_bug_on(sbi, dc->state != D_DONE);
dc->ref--;
if (!dc->ref) {
if (!dc->error)
len = dc->len;
__remove_discard_cmd(sbi, dc);
}
mutex_unlock(&dcc->cmd_lock);
return len;
}
static unsigned int __wait_discard_cmd_range(struct f2fs_sb_info *sbi,
struct discard_policy *dpolicy,
block_t start, block_t end)
{
struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
struct list_head *wait_list = (dpolicy->type == DPOLICY_FSTRIM) ?
&(dcc->fstrim_list) : &(dcc->wait_list);
struct discard_cmd *dc, *tmp;
bool need_wait;
unsigned int trimmed = 0;
next:
need_wait = false;
mutex_lock(&dcc->cmd_lock);
list_for_each_entry_safe(dc, tmp, wait_list, list) {
if (dc->lstart + dc->len <= start || end <= dc->lstart)
continue;
if (dc->len < dpolicy->granularity)
continue;
if (dc->state == D_DONE && !dc->ref) {
wait_for_completion_io(&dc->wait);
if (!dc->error)
trimmed += dc->len;
__remove_discard_cmd(sbi, dc);
} else {
dc->ref++;
need_wait = true;
break;
}
}
mutex_unlock(&dcc->cmd_lock);
if (need_wait) {
trimmed += __wait_one_discard_bio(sbi, dc);
goto next;
}
return trimmed;
}
static void __wait_all_discard_cmd(struct f2fs_sb_info *sbi,
struct discard_policy *dpolicy)
{
__wait_discard_cmd_range(sbi, dpolicy, 0, UINT_MAX);
}
/* This should be covered by global mutex, &sit_i->sentry_lock */
static void f2fs_wait_discard_bio(struct f2fs_sb_info *sbi, block_t blkaddr)
{
struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
struct discard_cmd *dc;
bool need_wait = false;
mutex_lock(&dcc->cmd_lock);
dc = (struct discard_cmd *)__lookup_rb_tree(&dcc->root, NULL, blkaddr);
if (dc) {
if (dc->state == D_PREP) {
__punch_discard_cmd(sbi, dc, blkaddr);
} else {
dc->ref++;
need_wait = true;
}
}
mutex_unlock(&dcc->cmd_lock);
if (need_wait)
__wait_one_discard_bio(sbi, dc);
}
void stop_discard_thread(struct f2fs_sb_info *sbi)
{
struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
if (dcc && dcc->f2fs_issue_discard) {
struct task_struct *discard_thread = dcc->f2fs_issue_discard;
dcc->f2fs_issue_discard = NULL;
kthread_stop(discard_thread);
}
}
/* This comes from f2fs_put_super */
bool f2fs_wait_discard_bios(struct f2fs_sb_info *sbi)
{
struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
struct discard_policy dpolicy;
bool dropped;
init_discard_policy(&dpolicy, DPOLICY_UMOUNT, dcc->discard_granularity);
__issue_discard_cmd(sbi, &dpolicy);
dropped = __drop_discard_cmd(sbi);
__wait_all_discard_cmd(sbi, &dpolicy);
return dropped;
}
static int issue_discard_thread(void *data)
{
struct f2fs_sb_info *sbi = data;
struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
wait_queue_head_t *q = &dcc->discard_wait_queue;
struct discard_policy dpolicy;
unsigned int wait_ms = DEF_MIN_DISCARD_ISSUE_TIME;
int issued;
set_freezable();
do {
init_discard_policy(&dpolicy, DPOLICY_BG,
dcc->discard_granularity);
wait_event_interruptible_timeout(*q,
kthread_should_stop() || freezing(current) ||
dcc->discard_wake,
msecs_to_jiffies(wait_ms));
if (try_to_freeze())
continue;
if (f2fs_readonly(sbi->sb))
continue;
if (kthread_should_stop())
return 0;
if (dcc->discard_wake)
dcc->discard_wake = 0;
if (sbi->gc_thread && sbi->gc_thread->gc_urgent)
init_discard_policy(&dpolicy, DPOLICY_FORCE, 1);
f2fs: make background threads of f2fs being aware of freezing When ->freeze_fs is called from lvm for doing snapshot, it needs to make sure there will be no more changes in filesystem's data, however, previously, background threads like GC thread wasn't aware of freezing, so in environment with active background threads, data of snapshot becomes unstable. This patch fixes this issue by adding sb_{start,end}_intwrite in below background threads: - GC thread - flush thread - discard thread Note that, don't use sb_start_intwrite() in gc_thread_func() due to: generic/241 reports below bug: ====================================================== WARNING: possible circular locking dependency detected 4.13.0-rc1+ #32 Tainted: G O ------------------------------------------------------ f2fs_gc-250:0/22186 is trying to acquire lock: (&sbi->gc_mutex){+.+...}, at: [<f8fa7f0b>] f2fs_sync_fs+0x7b/0x1b0 [f2fs] but task is already holding lock: (sb_internal#2){++++.-}, at: [<f8fb5609>] gc_thread_func+0x159/0x4a0 [f2fs] which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (sb_internal#2){++++.-}: __lock_acquire+0x405/0x7b0 lock_acquire+0xae/0x220 __sb_start_write+0x11d/0x1f0 f2fs_evict_inode+0x2d6/0x4e0 [f2fs] evict+0xa8/0x170 iput+0x1fb/0x2c0 f2fs_sync_inode_meta+0x3f/0xf0 [f2fs] write_checkpoint+0x1b1/0x750 [f2fs] f2fs_sync_fs+0x85/0x1b0 [f2fs] f2fs_do_sync_file.isra.24+0x137/0xa30 [f2fs] f2fs_sync_file+0x34/0x40 [f2fs] vfs_fsync_range+0x4a/0xa0 do_fsync+0x3c/0x60 SyS_fdatasync+0x15/0x20 do_fast_syscall_32+0xa1/0x1b0 entry_SYSENTER_32+0x4c/0x7b -> #1 (&sbi->cp_mutex){+.+...}: __lock_acquire+0x405/0x7b0 lock_acquire+0xae/0x220 __mutex_lock+0x4f/0x830 mutex_lock_nested+0x25/0x30 write_checkpoint+0x2f/0x750 [f2fs] f2fs_sync_fs+0x85/0x1b0 [f2fs] sync_filesystem+0x67/0x80 generic_shutdown_super+0x27/0x100 kill_block_super+0x22/0x50 kill_f2fs_super+0x3a/0x40 [f2fs] deactivate_locked_super+0x3d/0x70 deactivate_super+0x40/0x60 cleanup_mnt+0x39/0x70 __cleanup_mnt+0x10/0x20 task_work_run+0x69/0x80 exit_to_usermode_loop+0x57/0x92 do_fast_syscall_32+0x18c/0x1b0 entry_SYSENTER_32+0x4c/0x7b -> #0 (&sbi->gc_mutex){+.+...}: validate_chain.isra.36+0xc50/0xdb0 __lock_acquire+0x405/0x7b0 lock_acquire+0xae/0x220 __mutex_lock+0x4f/0x830 mutex_lock_nested+0x25/0x30 f2fs_sync_fs+0x7b/0x1b0 [f2fs] f2fs_balance_fs_bg+0xb9/0x200 [f2fs] gc_thread_func+0x302/0x4a0 [f2fs] kthread+0xe9/0x120 ret_from_fork+0x19/0x24 other info that might help us debug this: Chain exists of: &sbi->gc_mutex --> &sbi->cp_mutex --> sb_internal#2 Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(sb_internal#2); lock(&sbi->cp_mutex); lock(sb_internal#2); lock(&sbi->gc_mutex); *** DEADLOCK *** 1 lock held by f2fs_gc-250:0/22186: #0: (sb_internal#2){++++.-}, at: [<f8fb5609>] gc_thread_func+0x159/0x4a0 [f2fs] stack backtrace: CPU: 2 PID: 22186 Comm: f2fs_gc-250:0 Tainted: G O 4.13.0-rc1+ #32 Hardware name: innotek GmbH VirtualBox/VirtualBox, BIOS VirtualBox 12/01/2006 Call Trace: dump_stack+0x5f/0x92 print_circular_bug+0x1b3/0x1bd validate_chain.isra.36+0xc50/0xdb0 ? __this_cpu_preempt_check+0xf/0x20 __lock_acquire+0x405/0x7b0 lock_acquire+0xae/0x220 ? f2fs_sync_fs+0x7b/0x1b0 [f2fs] __mutex_lock+0x4f/0x830 ? f2fs_sync_fs+0x7b/0x1b0 [f2fs] mutex_lock_nested+0x25/0x30 ? f2fs_sync_fs+0x7b/0x1b0 [f2fs] f2fs_sync_fs+0x7b/0x1b0 [f2fs] f2fs_balance_fs_bg+0xb9/0x200 [f2fs] gc_thread_func+0x302/0x4a0 [f2fs] ? preempt_schedule_common+0x2f/0x4d ? f2fs_gc+0x540/0x540 [f2fs] kthread+0xe9/0x120 ? f2fs_gc+0x540/0x540 [f2fs] ? kthread_create_on_node+0x30/0x30 ret_from_fork+0x19/0x24 The deadlock occurs in below condition: GC Thread Thread B - sb_start_intwrite - f2fs_sync_file - f2fs_sync_fs - mutex_lock(&sbi->gc_mutex) - write_checkpoint - block_operations - f2fs_sync_inode_meta - iput - sb_start_intwrite - mutex_lock(&sbi->gc_mutex) Fix this by altering sb_start_intwrite to sb_start_write_trylock. Signed-off-by: Chao Yu <yuchao0@huawei.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2017-07-22 07:52:23 +07:00
sb_start_intwrite(sbi->sb);
issued = __issue_discard_cmd(sbi, &dpolicy);
if (issued) {
__wait_all_discard_cmd(sbi, &dpolicy);
wait_ms = dpolicy.min_interval;
} else {
wait_ms = dpolicy.max_interval;
}
f2fs: make background threads of f2fs being aware of freezing When ->freeze_fs is called from lvm for doing snapshot, it needs to make sure there will be no more changes in filesystem's data, however, previously, background threads like GC thread wasn't aware of freezing, so in environment with active background threads, data of snapshot becomes unstable. This patch fixes this issue by adding sb_{start,end}_intwrite in below background threads: - GC thread - flush thread - discard thread Note that, don't use sb_start_intwrite() in gc_thread_func() due to: generic/241 reports below bug: ====================================================== WARNING: possible circular locking dependency detected 4.13.0-rc1+ #32 Tainted: G O ------------------------------------------------------ f2fs_gc-250:0/22186 is trying to acquire lock: (&sbi->gc_mutex){+.+...}, at: [<f8fa7f0b>] f2fs_sync_fs+0x7b/0x1b0 [f2fs] but task is already holding lock: (sb_internal#2){++++.-}, at: [<f8fb5609>] gc_thread_func+0x159/0x4a0 [f2fs] which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (sb_internal#2){++++.-}: __lock_acquire+0x405/0x7b0 lock_acquire+0xae/0x220 __sb_start_write+0x11d/0x1f0 f2fs_evict_inode+0x2d6/0x4e0 [f2fs] evict+0xa8/0x170 iput+0x1fb/0x2c0 f2fs_sync_inode_meta+0x3f/0xf0 [f2fs] write_checkpoint+0x1b1/0x750 [f2fs] f2fs_sync_fs+0x85/0x1b0 [f2fs] f2fs_do_sync_file.isra.24+0x137/0xa30 [f2fs] f2fs_sync_file+0x34/0x40 [f2fs] vfs_fsync_range+0x4a/0xa0 do_fsync+0x3c/0x60 SyS_fdatasync+0x15/0x20 do_fast_syscall_32+0xa1/0x1b0 entry_SYSENTER_32+0x4c/0x7b -> #1 (&sbi->cp_mutex){+.+...}: __lock_acquire+0x405/0x7b0 lock_acquire+0xae/0x220 __mutex_lock+0x4f/0x830 mutex_lock_nested+0x25/0x30 write_checkpoint+0x2f/0x750 [f2fs] f2fs_sync_fs+0x85/0x1b0 [f2fs] sync_filesystem+0x67/0x80 generic_shutdown_super+0x27/0x100 kill_block_super+0x22/0x50 kill_f2fs_super+0x3a/0x40 [f2fs] deactivate_locked_super+0x3d/0x70 deactivate_super+0x40/0x60 cleanup_mnt+0x39/0x70 __cleanup_mnt+0x10/0x20 task_work_run+0x69/0x80 exit_to_usermode_loop+0x57/0x92 do_fast_syscall_32+0x18c/0x1b0 entry_SYSENTER_32+0x4c/0x7b -> #0 (&sbi->gc_mutex){+.+...}: validate_chain.isra.36+0xc50/0xdb0 __lock_acquire+0x405/0x7b0 lock_acquire+0xae/0x220 __mutex_lock+0x4f/0x830 mutex_lock_nested+0x25/0x30 f2fs_sync_fs+0x7b/0x1b0 [f2fs] f2fs_balance_fs_bg+0xb9/0x200 [f2fs] gc_thread_func+0x302/0x4a0 [f2fs] kthread+0xe9/0x120 ret_from_fork+0x19/0x24 other info that might help us debug this: Chain exists of: &sbi->gc_mutex --> &sbi->cp_mutex --> sb_internal#2 Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(sb_internal#2); lock(&sbi->cp_mutex); lock(sb_internal#2); lock(&sbi->gc_mutex); *** DEADLOCK *** 1 lock held by f2fs_gc-250:0/22186: #0: (sb_internal#2){++++.-}, at: [<f8fb5609>] gc_thread_func+0x159/0x4a0 [f2fs] stack backtrace: CPU: 2 PID: 22186 Comm: f2fs_gc-250:0 Tainted: G O 4.13.0-rc1+ #32 Hardware name: innotek GmbH VirtualBox/VirtualBox, BIOS VirtualBox 12/01/2006 Call Trace: dump_stack+0x5f/0x92 print_circular_bug+0x1b3/0x1bd validate_chain.isra.36+0xc50/0xdb0 ? __this_cpu_preempt_check+0xf/0x20 __lock_acquire+0x405/0x7b0 lock_acquire+0xae/0x220 ? f2fs_sync_fs+0x7b/0x1b0 [f2fs] __mutex_lock+0x4f/0x830 ? f2fs_sync_fs+0x7b/0x1b0 [f2fs] mutex_lock_nested+0x25/0x30 ? f2fs_sync_fs+0x7b/0x1b0 [f2fs] f2fs_sync_fs+0x7b/0x1b0 [f2fs] f2fs_balance_fs_bg+0xb9/0x200 [f2fs] gc_thread_func+0x302/0x4a0 [f2fs] ? preempt_schedule_common+0x2f/0x4d ? f2fs_gc+0x540/0x540 [f2fs] kthread+0xe9/0x120 ? f2fs_gc+0x540/0x540 [f2fs] ? kthread_create_on_node+0x30/0x30 ret_from_fork+0x19/0x24 The deadlock occurs in below condition: GC Thread Thread B - sb_start_intwrite - f2fs_sync_file - f2fs_sync_fs - mutex_lock(&sbi->gc_mutex) - write_checkpoint - block_operations - f2fs_sync_inode_meta - iput - sb_start_intwrite - mutex_lock(&sbi->gc_mutex) Fix this by altering sb_start_intwrite to sb_start_write_trylock. Signed-off-by: Chao Yu <yuchao0@huawei.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2017-07-22 07:52:23 +07:00
sb_end_intwrite(sbi->sb);
} while (!kthread_should_stop());
return 0;
}
#ifdef CONFIG_BLK_DEV_ZONED
static int __f2fs_issue_discard_zone(struct f2fs_sb_info *sbi,
struct block_device *bdev, block_t blkstart, block_t blklen)
{
sector_t sector, nr_sects;
block_t lblkstart = blkstart;
int devi = 0;
if (sbi->s_ndevs) {
devi = f2fs_target_device_index(sbi, blkstart);
blkstart -= FDEV(devi).start_blk;
}
/*
* We need to know the type of the zone: for conventional zones,
* use regular discard if the drive supports it. For sequential
* zones, reset the zone write pointer.
*/
switch (get_blkz_type(sbi, bdev, blkstart)) {
case BLK_ZONE_TYPE_CONVENTIONAL:
if (!blk_queue_discard(bdev_get_queue(bdev)))
return 0;
return __queue_discard_cmd(sbi, bdev, lblkstart, blklen);
case BLK_ZONE_TYPE_SEQWRITE_REQ:
case BLK_ZONE_TYPE_SEQWRITE_PREF:
sector = SECTOR_FROM_BLOCK(blkstart);
nr_sects = SECTOR_FROM_BLOCK(blklen);
if (sector & (bdev_zone_sectors(bdev) - 1) ||
nr_sects != bdev_zone_sectors(bdev)) {
f2fs_msg(sbi->sb, KERN_INFO,
"(%d) %s: Unaligned discard attempted (block %x + %x)",
devi, sbi->s_ndevs ? FDEV(devi).path: "",
blkstart, blklen);
return -EIO;
}
trace_f2fs_issue_reset_zone(bdev, blkstart);
return blkdev_reset_zones(bdev, sector,
nr_sects, GFP_NOFS);
default:
/* Unknown zone type: broken device ? */
return -EIO;
}
}
#endif
static int __issue_discard_async(struct f2fs_sb_info *sbi,
struct block_device *bdev, block_t blkstart, block_t blklen)
{
#ifdef CONFIG_BLK_DEV_ZONED
if (f2fs_sb_has_blkzoned(sbi->sb) &&
bdev_zoned_model(bdev) != BLK_ZONED_NONE)
return __f2fs_issue_discard_zone(sbi, bdev, blkstart, blklen);
#endif
return __queue_discard_cmd(sbi, bdev, blkstart, blklen);
}
f2fs: avoid to conduct roll-forward due to the remained garbage blocks The f2fs always scans the next chain of direct node blocks. But some garbage blocks are able to be remained due to no discard support or SSR triggers. This occasionally wreaks recovering wrong inodes that were used or BUG_ONs due to reallocating node ids as follows. When mount this f2fs image: http://linuxtesting.org/downloads/f2fs_fault_image.zip BUG_ON is triggered in f2fs driver (messages below are generated on kernel 3.13.2; for other kernels output is similar): kernel BUG at fs/f2fs/node.c:215! Call Trace: [<ffffffffa032ebad>] recover_inode_page+0x1fd/0x3e0 [f2fs] [<ffffffff811446e7>] ? __lock_page+0x67/0x70 [<ffffffff81089990>] ? autoremove_wake_function+0x50/0x50 [<ffffffffa0337788>] recover_fsync_data+0x1398/0x15d0 [f2fs] [<ffffffff812b9e5c>] ? selinux_d_instantiate+0x1c/0x20 [<ffffffff811cb20b>] ? d_instantiate+0x5b/0x80 [<ffffffffa0321044>] f2fs_fill_super+0xb04/0xbf0 [f2fs] [<ffffffff811b861e>] ? mount_bdev+0x7e/0x210 [<ffffffff811b8769>] mount_bdev+0x1c9/0x210 [<ffffffffa0320540>] ? validate_superblock+0x210/0x210 [f2fs] [<ffffffffa031cf8d>] f2fs_mount+0x1d/0x30 [f2fs] [<ffffffff811b9497>] mount_fs+0x47/0x1c0 [<ffffffff81166e00>] ? __alloc_percpu+0x10/0x20 [<ffffffff811d4032>] vfs_kern_mount+0x72/0x110 [<ffffffff811d6763>] do_mount+0x493/0x910 [<ffffffff811615cb>] ? strndup_user+0x5b/0x80 [<ffffffff811d6c70>] SyS_mount+0x90/0xe0 [<ffffffff8166f8d9>] system_call_fastpath+0x16/0x1b Found by Linux File System Verification project (linuxtesting.org). Reported-by: Andrey Tsyvarev <tsyvarev@ispras.ru> Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
2014-04-15 11:57:55 +07:00
static int f2fs_issue_discard(struct f2fs_sb_info *sbi,
block_t blkstart, block_t blklen)
{
sector_t start = blkstart, len = 0;
struct block_device *bdev;
struct seg_entry *se;
unsigned int offset;
block_t i;
int err = 0;
bdev = f2fs_target_device(sbi, blkstart, NULL);
for (i = blkstart; i < blkstart + blklen; i++, len++) {
if (i != start) {
struct block_device *bdev2 =
f2fs_target_device(sbi, i, NULL);
if (bdev2 != bdev) {
err = __issue_discard_async(sbi, bdev,
start, len);
if (err)
return err;
bdev = bdev2;
start = i;
len = 0;
}
}
se = get_seg_entry(sbi, GET_SEGNO(sbi, i));
offset = GET_BLKOFF_FROM_SEG0(sbi, i);
if (!f2fs_test_and_set_bit(offset, se->discard_map))
sbi->discard_blks--;
}
if (len)
err = __issue_discard_async(sbi, bdev, start, len);
return err;
f2fs: avoid to conduct roll-forward due to the remained garbage blocks The f2fs always scans the next chain of direct node blocks. But some garbage blocks are able to be remained due to no discard support or SSR triggers. This occasionally wreaks recovering wrong inodes that were used or BUG_ONs due to reallocating node ids as follows. When mount this f2fs image: http://linuxtesting.org/downloads/f2fs_fault_image.zip BUG_ON is triggered in f2fs driver (messages below are generated on kernel 3.13.2; for other kernels output is similar): kernel BUG at fs/f2fs/node.c:215! Call Trace: [<ffffffffa032ebad>] recover_inode_page+0x1fd/0x3e0 [f2fs] [<ffffffff811446e7>] ? __lock_page+0x67/0x70 [<ffffffff81089990>] ? autoremove_wake_function+0x50/0x50 [<ffffffffa0337788>] recover_fsync_data+0x1398/0x15d0 [f2fs] [<ffffffff812b9e5c>] ? selinux_d_instantiate+0x1c/0x20 [<ffffffff811cb20b>] ? d_instantiate+0x5b/0x80 [<ffffffffa0321044>] f2fs_fill_super+0xb04/0xbf0 [f2fs] [<ffffffff811b861e>] ? mount_bdev+0x7e/0x210 [<ffffffff811b8769>] mount_bdev+0x1c9/0x210 [<ffffffffa0320540>] ? validate_superblock+0x210/0x210 [f2fs] [<ffffffffa031cf8d>] f2fs_mount+0x1d/0x30 [f2fs] [<ffffffff811b9497>] mount_fs+0x47/0x1c0 [<ffffffff81166e00>] ? __alloc_percpu+0x10/0x20 [<ffffffff811d4032>] vfs_kern_mount+0x72/0x110 [<ffffffff811d6763>] do_mount+0x493/0x910 [<ffffffff811615cb>] ? strndup_user+0x5b/0x80 [<ffffffff811d6c70>] SyS_mount+0x90/0xe0 [<ffffffff8166f8d9>] system_call_fastpath+0x16/0x1b Found by Linux File System Verification project (linuxtesting.org). Reported-by: Andrey Tsyvarev <tsyvarev@ispras.ru> Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
2014-04-15 11:57:55 +07:00
}
static bool add_discard_addrs(struct f2fs_sb_info *sbi, struct cp_control *cpc,
bool check_only)
{
int entries = SIT_VBLOCK_MAP_SIZE / sizeof(unsigned long);
int max_blocks = sbi->blocks_per_seg;
struct seg_entry *se = get_seg_entry(sbi, cpc->trim_start);
unsigned long *cur_map = (unsigned long *)se->cur_valid_map;
unsigned long *ckpt_map = (unsigned long *)se->ckpt_valid_map;
unsigned long *discard_map = (unsigned long *)se->discard_map;
unsigned long *dmap = SIT_I(sbi)->tmp_map;
unsigned int start = 0, end = -1;
bool force = (cpc->reason & CP_DISCARD);
struct discard_entry *de = NULL;
struct list_head *head = &SM_I(sbi)->dcc_info->entry_list;
int i;
if (se->valid_blocks == max_blocks || !f2fs_discard_en(sbi))
return false;
if (!force) {
if (!test_opt(sbi, DISCARD) || !se->valid_blocks ||
SM_I(sbi)->dcc_info->nr_discards >=
SM_I(sbi)->dcc_info->max_discards)
return false;
}
/* SIT_VBLOCK_MAP_SIZE should be multiple of sizeof(unsigned long) */
for (i = 0; i < entries; i++)
dmap[i] = force ? ~ckpt_map[i] & ~discard_map[i] :
(cur_map[i] ^ ckpt_map[i]) & ckpt_map[i];
while (force || SM_I(sbi)->dcc_info->nr_discards <=
SM_I(sbi)->dcc_info->max_discards) {
start = __find_rev_next_bit(dmap, max_blocks, end + 1);
if (start >= max_blocks)
break;
end = __find_rev_next_zero_bit(dmap, max_blocks, start + 1);
if (force && start && end != max_blocks
&& (end - start) < cpc->trim_minlen)
continue;
if (check_only)
return true;
if (!de) {
de = f2fs_kmem_cache_alloc(discard_entry_slab,
GFP_F2FS_ZERO);
de->start_blkaddr = START_BLOCK(sbi, cpc->trim_start);
list_add_tail(&de->list, head);
}
for (i = start; i < end; i++)
__set_bit_le(i, (void *)de->discard_map);
SM_I(sbi)->dcc_info->nr_discards += end - start;
}
return false;
}
void release_discard_addrs(struct f2fs_sb_info *sbi)
{
struct list_head *head = &(SM_I(sbi)->dcc_info->entry_list);
struct discard_entry *entry, *this;
/* drop caches */
list_for_each_entry_safe(entry, this, head, list) {
list_del(&entry->list);
kmem_cache_free(discard_entry_slab, entry);
}
}
/*
* Should call clear_prefree_segments after checkpoint is done.
*/
static void set_prefree_as_free_segments(struct f2fs_sb_info *sbi)
{
struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
unsigned int segno;
mutex_lock(&dirty_i->seglist_lock);
for_each_set_bit(segno, dirty_i->dirty_segmap[PRE], MAIN_SEGS(sbi))
__set_test_and_free(sbi, segno);
mutex_unlock(&dirty_i->seglist_lock);
}
void clear_prefree_segments(struct f2fs_sb_info *sbi, struct cp_control *cpc)
{
struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
struct list_head *head = &dcc->entry_list;
struct discard_entry *entry, *this;
struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
unsigned long *prefree_map = dirty_i->dirty_segmap[PRE];
unsigned int start = 0, end = -1;
unsigned int secno, start_segno;
bool force = (cpc->reason & CP_DISCARD);
mutex_lock(&dirty_i->seglist_lock);
while (1) {
int i;
start = find_next_bit(prefree_map, MAIN_SEGS(sbi), end + 1);
if (start >= MAIN_SEGS(sbi))
break;
end = find_next_zero_bit(prefree_map, MAIN_SEGS(sbi),
start + 1);
for (i = start; i < end; i++)
clear_bit(i, prefree_map);
dirty_i->nr_dirty[PRE] -= end - start;
if (!test_opt(sbi, DISCARD))
continue;
if (force && start >= cpc->trim_start &&
(end - 1) <= cpc->trim_end)
continue;
if (!test_opt(sbi, LFS) || sbi->segs_per_sec == 1) {
f2fs_issue_discard(sbi, START_BLOCK(sbi, start),
(end - start) << sbi->log_blocks_per_seg);
continue;
}
next:
secno = GET_SEC_FROM_SEG(sbi, start);
start_segno = GET_SEG_FROM_SEC(sbi, secno);
if (!IS_CURSEC(sbi, secno) &&
!get_valid_blocks(sbi, start, true))
f2fs_issue_discard(sbi, START_BLOCK(sbi, start_segno),
sbi->segs_per_sec << sbi->log_blocks_per_seg);
start = start_segno + sbi->segs_per_sec;
if (start < end)
goto next;
else
end = start - 1;
}
mutex_unlock(&dirty_i->seglist_lock);
/* send small discards */
list_for_each_entry_safe(entry, this, head, list) {
unsigned int cur_pos = 0, next_pos, len, total_len = 0;
bool is_valid = test_bit_le(0, entry->discard_map);
find_next:
if (is_valid) {
next_pos = find_next_zero_bit_le(entry->discard_map,
sbi->blocks_per_seg, cur_pos);
len = next_pos - cur_pos;
if (f2fs_sb_has_blkzoned(sbi->sb) ||
(force && len < cpc->trim_minlen))
goto skip;
f2fs_issue_discard(sbi, entry->start_blkaddr + cur_pos,
len);
total_len += len;
} else {
next_pos = find_next_bit_le(entry->discard_map,
sbi->blocks_per_seg, cur_pos);
}
skip:
cur_pos = next_pos;
is_valid = !is_valid;
if (cur_pos < sbi->blocks_per_seg)
goto find_next;
list_del(&entry->list);
dcc->nr_discards -= total_len;
kmem_cache_free(discard_entry_slab, entry);
}
wake_up_discard_thread(sbi, false);
}
void init_discard_policy(struct discard_policy *dpolicy,
int discard_type, unsigned int granularity)
{
/* common policy */
dpolicy->type = discard_type;
dpolicy->sync = true;
dpolicy->granularity = granularity;
dpolicy->max_requests = DEF_MAX_DISCARD_REQUEST;
dpolicy->io_aware_gran = MAX_PLIST_NUM;
if (discard_type == DPOLICY_BG) {
dpolicy->min_interval = DEF_MIN_DISCARD_ISSUE_TIME;
dpolicy->max_interval = DEF_MAX_DISCARD_ISSUE_TIME;
dpolicy->io_aware = true;
} else if (discard_type == DPOLICY_FORCE) {
dpolicy->min_interval = DEF_MIN_DISCARD_ISSUE_TIME;
dpolicy->max_interval = DEF_MAX_DISCARD_ISSUE_TIME;
dpolicy->io_aware = false;
} else if (discard_type == DPOLICY_FSTRIM) {
dpolicy->io_aware = false;
} else if (discard_type == DPOLICY_UMOUNT) {
dpolicy->io_aware = false;
}
}
static int create_discard_cmd_control(struct f2fs_sb_info *sbi)
{
dev_t dev = sbi->sb->s_bdev->bd_dev;
struct discard_cmd_control *dcc;
int err = 0, i;
if (SM_I(sbi)->dcc_info) {
dcc = SM_I(sbi)->dcc_info;
goto init_thread;
}
dcc = f2fs_kzalloc(sbi, sizeof(struct discard_cmd_control), GFP_KERNEL);
if (!dcc)
return -ENOMEM;
dcc->discard_granularity = DEFAULT_DISCARD_GRANULARITY;
INIT_LIST_HEAD(&dcc->entry_list);
for (i = 0; i < MAX_PLIST_NUM; i++)
INIT_LIST_HEAD(&dcc->pend_list[i]);
INIT_LIST_HEAD(&dcc->wait_list);
INIT_LIST_HEAD(&dcc->fstrim_list);
mutex_init(&dcc->cmd_lock);
atomic_set(&dcc->issued_discard, 0);
atomic_set(&dcc->issing_discard, 0);
atomic_set(&dcc->discard_cmd_cnt, 0);
dcc->nr_discards = 0;
dcc->max_discards = MAIN_SEGS(sbi) << sbi->log_blocks_per_seg;
dcc->undiscard_blks = 0;
dcc->root = RB_ROOT;
init_waitqueue_head(&dcc->discard_wait_queue);
SM_I(sbi)->dcc_info = dcc;
init_thread:
dcc->f2fs_issue_discard = kthread_run(issue_discard_thread, sbi,
"f2fs_discard-%u:%u", MAJOR(dev), MINOR(dev));
if (IS_ERR(dcc->f2fs_issue_discard)) {
err = PTR_ERR(dcc->f2fs_issue_discard);
kfree(dcc);
SM_I(sbi)->dcc_info = NULL;
return err;
}
return err;
}
static void destroy_discard_cmd_control(struct f2fs_sb_info *sbi)
{
struct discard_cmd_control *dcc = SM_I(sbi)->dcc_info;
if (!dcc)
return;
stop_discard_thread(sbi);
kfree(dcc);
SM_I(sbi)->dcc_info = NULL;
}
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
static bool __mark_sit_entry_dirty(struct f2fs_sb_info *sbi, unsigned int segno)
{
struct sit_info *sit_i = SIT_I(sbi);
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
if (!__test_and_set_bit(segno, sit_i->dirty_sentries_bitmap)) {
sit_i->dirty_sentries++;
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
return false;
}
return true;
}
static void __set_sit_entry_type(struct f2fs_sb_info *sbi, int type,
unsigned int segno, int modified)
{
struct seg_entry *se = get_seg_entry(sbi, segno);
se->type = type;
if (modified)
__mark_sit_entry_dirty(sbi, segno);
}
static void update_sit_entry(struct f2fs_sb_info *sbi, block_t blkaddr, int del)
{
struct seg_entry *se;
unsigned int segno, offset;
long int new_vblocks;
bool exist;
#ifdef CONFIG_F2FS_CHECK_FS
bool mir_exist;
#endif
segno = GET_SEGNO(sbi, blkaddr);
se = get_seg_entry(sbi, segno);
new_vblocks = se->valid_blocks + del;
offset = GET_BLKOFF_FROM_SEG0(sbi, blkaddr);
f2fs_bug_on(sbi, (new_vblocks >> (sizeof(unsigned short) << 3) ||
(new_vblocks > sbi->blocks_per_seg)));
se->valid_blocks = new_vblocks;
se->mtime = get_mtime(sbi);
SIT_I(sbi)->max_mtime = se->mtime;
/* Update valid block bitmap */
if (del > 0) {
exist = f2fs_test_and_set_bit(offset, se->cur_valid_map);
#ifdef CONFIG_F2FS_CHECK_FS
mir_exist = f2fs_test_and_set_bit(offset,
se->cur_valid_map_mir);
if (unlikely(exist != mir_exist)) {
f2fs_msg(sbi->sb, KERN_ERR, "Inconsistent error "
"when setting bitmap, blk:%u, old bit:%d",
blkaddr, exist);
f2fs_bug_on(sbi, 1);
}
#endif
if (unlikely(exist)) {
f2fs_msg(sbi->sb, KERN_ERR,
"Bitmap was wrongly set, blk:%u", blkaddr);
f2fs_bug_on(sbi, 1);
se->valid_blocks--;
del = 0;
}
if (f2fs_discard_en(sbi) &&
!f2fs_test_and_set_bit(offset, se->discard_map))
sbi->discard_blks--;
/* don't overwrite by SSR to keep node chain */
if (IS_NODESEG(se->type)) {
if (!f2fs_test_and_set_bit(offset, se->ckpt_valid_map))
se->ckpt_valid_blocks++;
}
} else {
exist = f2fs_test_and_clear_bit(offset, se->cur_valid_map);
#ifdef CONFIG_F2FS_CHECK_FS
mir_exist = f2fs_test_and_clear_bit(offset,
se->cur_valid_map_mir);
if (unlikely(exist != mir_exist)) {
f2fs_msg(sbi->sb, KERN_ERR, "Inconsistent error "
"when clearing bitmap, blk:%u, old bit:%d",
blkaddr, exist);
f2fs_bug_on(sbi, 1);
}
#endif
if (unlikely(!exist)) {
f2fs_msg(sbi->sb, KERN_ERR,
"Bitmap was wrongly cleared, blk:%u", blkaddr);
f2fs_bug_on(sbi, 1);
se->valid_blocks++;
del = 0;
}
if (f2fs_discard_en(sbi) &&
f2fs_test_and_clear_bit(offset, se->discard_map))
sbi->discard_blks++;
}
if (!f2fs_test_bit(offset, se->ckpt_valid_map))
se->ckpt_valid_blocks += del;
__mark_sit_entry_dirty(sbi, segno);
/* update total number of valid blocks to be written in ckpt area */
SIT_I(sbi)->written_valid_blocks += del;
if (sbi->segs_per_sec > 1)
get_sec_entry(sbi, segno)->valid_blocks += del;
}
void invalidate_blocks(struct f2fs_sb_info *sbi, block_t addr)
{
unsigned int segno = GET_SEGNO(sbi, addr);
struct sit_info *sit_i = SIT_I(sbi);
f2fs_bug_on(sbi, addr == NULL_ADDR);
if (addr == NEW_ADDR)
return;
/* add it into sit main buffer */
down_write(&sit_i->sentry_lock);
update_sit_entry(sbi, addr, -1);
/* add it into dirty seglist */
locate_dirty_segment(sbi, segno);
up_write(&sit_i->sentry_lock);
}
bool is_checkpointed_data(struct f2fs_sb_info *sbi, block_t blkaddr)
{
struct sit_info *sit_i = SIT_I(sbi);
unsigned int segno, offset;
struct seg_entry *se;
bool is_cp = false;
if (blkaddr == NEW_ADDR || blkaddr == NULL_ADDR)
return true;
down_read(&sit_i->sentry_lock);
segno = GET_SEGNO(sbi, blkaddr);
se = get_seg_entry(sbi, segno);
offset = GET_BLKOFF_FROM_SEG0(sbi, blkaddr);
if (f2fs_test_bit(offset, se->ckpt_valid_map))
is_cp = true;
up_read(&sit_i->sentry_lock);
return is_cp;
}
/*
* This function should be resided under the curseg_mutex lock
*/
static void __add_sum_entry(struct f2fs_sb_info *sbi, int type,
struct f2fs_summary *sum)
{
struct curseg_info *curseg = CURSEG_I(sbi, type);
void *addr = curseg->sum_blk;
addr += curseg->next_blkoff * sizeof(struct f2fs_summary);
memcpy(addr, sum, sizeof(struct f2fs_summary));
}
/*
* Calculate the number of current summary pages for writing
*/
int npages_for_summary_flush(struct f2fs_sb_info *sbi, bool for_ra)
{
int valid_sum_count = 0;
int i, sum_in_page;
for (i = CURSEG_HOT_DATA; i <= CURSEG_COLD_DATA; i++) {
if (sbi->ckpt->alloc_type[i] == SSR)
valid_sum_count += sbi->blocks_per_seg;
else {
if (for_ra)
valid_sum_count += le16_to_cpu(
F2FS_CKPT(sbi)->cur_data_blkoff[i]);
else
valid_sum_count += curseg_blkoff(sbi, i);
}
}
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 19:29:47 +07:00
sum_in_page = (PAGE_SIZE - 2 * SUM_JOURNAL_SIZE -
SUM_FOOTER_SIZE) / SUMMARY_SIZE;
if (valid_sum_count <= sum_in_page)
return 1;
else if ((valid_sum_count - sum_in_page) <=
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 19:29:47 +07:00
(PAGE_SIZE - SUM_FOOTER_SIZE) / SUMMARY_SIZE)
return 2;
return 3;
}
/*
* Caller should put this summary page
*/
struct page *get_sum_page(struct f2fs_sb_info *sbi, unsigned int segno)
{
return get_meta_page(sbi, GET_SUM_BLOCK(sbi, segno));
}
void update_meta_page(struct f2fs_sb_info *sbi, void *src, block_t blk_addr)
{
struct page *page = grab_meta_page(sbi, blk_addr);
memcpy(page_address(page), src, PAGE_SIZE);
set_page_dirty(page);
f2fs_put_page(page, 1);
}
static void write_sum_page(struct f2fs_sb_info *sbi,
struct f2fs_summary_block *sum_blk, block_t blk_addr)
{
update_meta_page(sbi, (void *)sum_blk, blk_addr);
}
static void write_current_sum_page(struct f2fs_sb_info *sbi,
int type, block_t blk_addr)
{
struct curseg_info *curseg = CURSEG_I(sbi, type);
struct page *page = grab_meta_page(sbi, blk_addr);
struct f2fs_summary_block *src = curseg->sum_blk;
struct f2fs_summary_block *dst;
dst = (struct f2fs_summary_block *)page_address(page);
mutex_lock(&curseg->curseg_mutex);
down_read(&curseg->journal_rwsem);
memcpy(&dst->journal, curseg->journal, SUM_JOURNAL_SIZE);
up_read(&curseg->journal_rwsem);
memcpy(dst->entries, src->entries, SUM_ENTRY_SIZE);
memcpy(&dst->footer, &src->footer, SUM_FOOTER_SIZE);
mutex_unlock(&curseg->curseg_mutex);
set_page_dirty(page);
f2fs_put_page(page, 1);
}
static int is_next_segment_free(struct f2fs_sb_info *sbi, int type)
{
struct curseg_info *curseg = CURSEG_I(sbi, type);
unsigned int segno = curseg->segno + 1;
struct free_segmap_info *free_i = FREE_I(sbi);
if (segno < MAIN_SEGS(sbi) && segno % sbi->segs_per_sec)
return !test_bit(segno, free_i->free_segmap);
return 0;
}
/*
* Find a new segment from the free segments bitmap to right order
* This function should be returned with success, otherwise BUG
*/
static void get_new_segment(struct f2fs_sb_info *sbi,
unsigned int *newseg, bool new_sec, int dir)
{
struct free_segmap_info *free_i = FREE_I(sbi);
unsigned int segno, secno, zoneno;
unsigned int total_zones = MAIN_SECS(sbi) / sbi->secs_per_zone;
unsigned int hint = GET_SEC_FROM_SEG(sbi, *newseg);
unsigned int old_zoneno = GET_ZONE_FROM_SEG(sbi, *newseg);
unsigned int left_start = hint;
bool init = true;
int go_left = 0;
int i;
spin_lock(&free_i->segmap_lock);
if (!new_sec && ((*newseg + 1) % sbi->segs_per_sec)) {
segno = find_next_zero_bit(free_i->free_segmap,
GET_SEG_FROM_SEC(sbi, hint + 1), *newseg + 1);
if (segno < GET_SEG_FROM_SEC(sbi, hint + 1))
goto got_it;
}
find_other_zone:
secno = find_next_zero_bit(free_i->free_secmap, MAIN_SECS(sbi), hint);
if (secno >= MAIN_SECS(sbi)) {
if (dir == ALLOC_RIGHT) {
secno = find_next_zero_bit(free_i->free_secmap,
MAIN_SECS(sbi), 0);
f2fs_bug_on(sbi, secno >= MAIN_SECS(sbi));
} else {
go_left = 1;
left_start = hint - 1;
}
}
if (go_left == 0)
goto skip_left;
while (test_bit(left_start, free_i->free_secmap)) {
if (left_start > 0) {
left_start--;
continue;
}
left_start = find_next_zero_bit(free_i->free_secmap,
MAIN_SECS(sbi), 0);
f2fs_bug_on(sbi, left_start >= MAIN_SECS(sbi));
break;
}
secno = left_start;
skip_left:
segno = GET_SEG_FROM_SEC(sbi, secno);
zoneno = GET_ZONE_FROM_SEC(sbi, secno);
/* give up on finding another zone */
if (!init)
goto got_it;
if (sbi->secs_per_zone == 1)
goto got_it;
if (zoneno == old_zoneno)
goto got_it;
if (dir == ALLOC_LEFT) {
if (!go_left && zoneno + 1 >= total_zones)
goto got_it;
if (go_left && zoneno == 0)
goto got_it;
}
for (i = 0; i < NR_CURSEG_TYPE; i++)
if (CURSEG_I(sbi, i)->zone == zoneno)
break;
if (i < NR_CURSEG_TYPE) {
/* zone is in user, try another */
if (go_left)
hint = zoneno * sbi->secs_per_zone - 1;
else if (zoneno + 1 >= total_zones)
hint = 0;
else
hint = (zoneno + 1) * sbi->secs_per_zone;
init = false;
goto find_other_zone;
}
got_it:
/* set it as dirty segment in free segmap */
f2fs_bug_on(sbi, test_bit(segno, free_i->free_segmap));
__set_inuse(sbi, segno);
*newseg = segno;
spin_unlock(&free_i->segmap_lock);
}
static void reset_curseg(struct f2fs_sb_info *sbi, int type, int modified)
{
struct curseg_info *curseg = CURSEG_I(sbi, type);
struct summary_footer *sum_footer;
curseg->segno = curseg->next_segno;
curseg->zone = GET_ZONE_FROM_SEG(sbi, curseg->segno);
curseg->next_blkoff = 0;
curseg->next_segno = NULL_SEGNO;
sum_footer = &(curseg->sum_blk->footer);
memset(sum_footer, 0, sizeof(struct summary_footer));
if (IS_DATASEG(type))
SET_SUM_TYPE(sum_footer, SUM_TYPE_DATA);
if (IS_NODESEG(type))
SET_SUM_TYPE(sum_footer, SUM_TYPE_NODE);
__set_sit_entry_type(sbi, type, curseg->segno, modified);
}
static unsigned int __get_next_segno(struct f2fs_sb_info *sbi, int type)
{
/* if segs_per_sec is large than 1, we need to keep original policy. */
if (sbi->segs_per_sec != 1)
return CURSEG_I(sbi, type)->segno;
if (test_opt(sbi, NOHEAP) &&
(type == CURSEG_HOT_DATA || IS_NODESEG(type)))
return 0;
if (SIT_I(sbi)->last_victim[ALLOC_NEXT])
return SIT_I(sbi)->last_victim[ALLOC_NEXT];
/* find segments from 0 to reuse freed segments */
if (F2FS_OPTION(sbi).alloc_mode == ALLOC_MODE_REUSE)
return 0;
return CURSEG_I(sbi, type)->segno;
}
/*
* Allocate a current working segment.
* This function always allocates a free segment in LFS manner.
*/
static void new_curseg(struct f2fs_sb_info *sbi, int type, bool new_sec)
{
struct curseg_info *curseg = CURSEG_I(sbi, type);
unsigned int segno = curseg->segno;
int dir = ALLOC_LEFT;
write_sum_page(sbi, curseg->sum_blk,
GET_SUM_BLOCK(sbi, segno));
if (type == CURSEG_WARM_DATA || type == CURSEG_COLD_DATA)
dir = ALLOC_RIGHT;
if (test_opt(sbi, NOHEAP))
dir = ALLOC_RIGHT;
segno = __get_next_segno(sbi, type);
get_new_segment(sbi, &segno, new_sec, dir);
curseg->next_segno = segno;
reset_curseg(sbi, type, 1);
curseg->alloc_type = LFS;
}
static void __next_free_blkoff(struct f2fs_sb_info *sbi,
struct curseg_info *seg, block_t start)
{
struct seg_entry *se = get_seg_entry(sbi, seg->segno);
int entries = SIT_VBLOCK_MAP_SIZE / sizeof(unsigned long);
unsigned long *target_map = SIT_I(sbi)->tmp_map;
unsigned long *ckpt_map = (unsigned long *)se->ckpt_valid_map;
unsigned long *cur_map = (unsigned long *)se->cur_valid_map;
int i, pos;
for (i = 0; i < entries; i++)
target_map[i] = ckpt_map[i] | cur_map[i];
pos = __find_rev_next_zero_bit(target_map, sbi->blocks_per_seg, start);
seg->next_blkoff = pos;
}
/*
* If a segment is written by LFS manner, next block offset is just obtained
* by increasing the current block offset. However, if a segment is written by
* SSR manner, next block offset obtained by calling __next_free_blkoff
*/
static void __refresh_next_blkoff(struct f2fs_sb_info *sbi,
struct curseg_info *seg)
{
if (seg->alloc_type == SSR)
__next_free_blkoff(sbi, seg, seg->next_blkoff + 1);
else
seg->next_blkoff++;
}
/*
* This function always allocates a used segment(from dirty seglist) by SSR
* manner, so it should recover the existing segment information of valid blocks
*/
static void change_curseg(struct f2fs_sb_info *sbi, int type)
{
struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
struct curseg_info *curseg = CURSEG_I(sbi, type);
unsigned int new_segno = curseg->next_segno;
struct f2fs_summary_block *sum_node;
struct page *sum_page;
write_sum_page(sbi, curseg->sum_blk,
GET_SUM_BLOCK(sbi, curseg->segno));
__set_test_and_inuse(sbi, new_segno);
mutex_lock(&dirty_i->seglist_lock);
__remove_dirty_segment(sbi, new_segno, PRE);
__remove_dirty_segment(sbi, new_segno, DIRTY);
mutex_unlock(&dirty_i->seglist_lock);
reset_curseg(sbi, type, 1);
curseg->alloc_type = SSR;
__next_free_blkoff(sbi, curseg, 0);
sum_page = get_sum_page(sbi, new_segno);
sum_node = (struct f2fs_summary_block *)page_address(sum_page);
memcpy(curseg->sum_blk, sum_node, SUM_ENTRY_SIZE);
f2fs_put_page(sum_page, 1);
}
static int get_ssr_segment(struct f2fs_sb_info *sbi, int type)
{
struct curseg_info *curseg = CURSEG_I(sbi, type);
const struct victim_selection *v_ops = DIRTY_I(sbi)->v_ops;
unsigned segno = NULL_SEGNO;
int i, cnt;
bool reversed = false;
/* need_SSR() already forces to do this */
if (v_ops->get_victim(sbi, &segno, BG_GC, type, SSR)) {
curseg->next_segno = segno;
return 1;
}
/* For node segments, let's do SSR more intensively */
if (IS_NODESEG(type)) {
if (type >= CURSEG_WARM_NODE) {
reversed = true;
i = CURSEG_COLD_NODE;
} else {
i = CURSEG_HOT_NODE;
}
cnt = NR_CURSEG_NODE_TYPE;
} else {
if (type >= CURSEG_WARM_DATA) {
reversed = true;
i = CURSEG_COLD_DATA;
} else {
i = CURSEG_HOT_DATA;
}
cnt = NR_CURSEG_DATA_TYPE;
}
for (; cnt-- > 0; reversed ? i-- : i++) {
if (i == type)
continue;
if (v_ops->get_victim(sbi, &segno, BG_GC, i, SSR)) {
curseg->next_segno = segno;
return 1;
}
}
return 0;
}
/*
* flush out current segment and replace it with new segment
* This function should be returned with success, otherwise BUG
*/
static void allocate_segment_by_default(struct f2fs_sb_info *sbi,
int type, bool force)
{
struct curseg_info *curseg = CURSEG_I(sbi, type);
if (force)
new_curseg(sbi, type, true);
else if (!is_set_ckpt_flags(sbi, CP_CRC_RECOVERY_FLAG) &&
type == CURSEG_WARM_NODE)
new_curseg(sbi, type, false);
else if (curseg->alloc_type == LFS && is_next_segment_free(sbi, type))
new_curseg(sbi, type, false);
else if (need_SSR(sbi) && get_ssr_segment(sbi, type))
change_curseg(sbi, type);
else
new_curseg(sbi, type, false);
stat_inc_seg_type(sbi, curseg);
}
void allocate_new_segments(struct f2fs_sb_info *sbi)
{
struct curseg_info *curseg;
unsigned int old_segno;
int i;
down_write(&SIT_I(sbi)->sentry_lock);
for (i = CURSEG_HOT_DATA; i <= CURSEG_COLD_DATA; i++) {
curseg = CURSEG_I(sbi, i);
old_segno = curseg->segno;
SIT_I(sbi)->s_ops->allocate_segment(sbi, i, true);
locate_dirty_segment(sbi, old_segno);
}
up_write(&SIT_I(sbi)->sentry_lock);
}
static const struct segment_allocation default_salloc_ops = {
.allocate_segment = allocate_segment_by_default,
};
bool exist_trim_candidates(struct f2fs_sb_info *sbi, struct cp_control *cpc)
{
__u64 trim_start = cpc->trim_start;
bool has_candidate = false;
down_write(&SIT_I(sbi)->sentry_lock);
for (; cpc->trim_start <= cpc->trim_end; cpc->trim_start++) {
if (add_discard_addrs(sbi, cpc, true)) {
has_candidate = true;
break;
}
}
up_write(&SIT_I(sbi)->sentry_lock);
cpc->trim_start = trim_start;
return has_candidate;
}
int f2fs_trim_fs(struct f2fs_sb_info *sbi, struct fstrim_range *range)
{
__u64 start = F2FS_BYTES_TO_BLK(range->start);
__u64 end = start + F2FS_BYTES_TO_BLK(range->len) - 1;
unsigned int start_segno, end_segno, cur_segno;
block_t start_block, end_block;
struct cp_control cpc;
struct discard_policy dpolicy;
unsigned long long trimmed = 0;
int err = 0;
if (start >= MAX_BLKADDR(sbi) || range->len < sbi->blocksize)
return -EINVAL;
if (end <= MAIN_BLKADDR(sbi))
goto out;
if (is_sbi_flag_set(sbi, SBI_NEED_FSCK)) {
f2fs_msg(sbi->sb, KERN_WARNING,
"Found FS corruption, run fsck to fix.");
goto out;
}
/* start/end segment number in main_area */
start_segno = (start <= MAIN_BLKADDR(sbi)) ? 0 : GET_SEGNO(sbi, start);
end_segno = (end >= MAX_BLKADDR(sbi)) ? MAIN_SEGS(sbi) - 1 :
GET_SEGNO(sbi, end);
cpc.reason = CP_DISCARD;
cpc.trim_minlen = max_t(__u64, 1, F2FS_BYTES_TO_BLK(range->minlen));
/* do checkpoint to issue discard commands safely */
for (cur_segno = start_segno; cur_segno <= end_segno;
cur_segno = cpc.trim_end + 1) {
cpc.trim_start = cur_segno;
if (sbi->discard_blks == 0)
break;
else if (sbi->discard_blks < BATCHED_TRIM_BLOCKS(sbi))
cpc.trim_end = end_segno;
else
cpc.trim_end = min_t(unsigned int,
rounddown(cur_segno +
BATCHED_TRIM_SEGMENTS(sbi),
sbi->segs_per_sec) - 1, end_segno);
mutex_lock(&sbi->gc_mutex);
err = write_checkpoint(sbi, &cpc);
mutex_unlock(&sbi->gc_mutex);
if (err)
break;
schedule();
}
start_block = START_BLOCK(sbi, start_segno);
end_block = START_BLOCK(sbi, min(cur_segno, end_segno) + 1);
init_discard_policy(&dpolicy, DPOLICY_FSTRIM, cpc.trim_minlen);
__issue_discard_cmd_range(sbi, &dpolicy, start_block, end_block);
trimmed = __wait_discard_cmd_range(sbi, &dpolicy,
start_block, end_block);
out:
range->len = F2FS_BLK_TO_BYTES(trimmed);
return err;
}
static bool __has_curseg_space(struct f2fs_sb_info *sbi, int type)
{
struct curseg_info *curseg = CURSEG_I(sbi, type);
if (curseg->next_blkoff < sbi->blocks_per_seg)
return true;
return false;
}
int rw_hint_to_seg_type(enum rw_hint hint)
{
switch (hint) {
case WRITE_LIFE_SHORT:
return CURSEG_HOT_DATA;
case WRITE_LIFE_EXTREME:
return CURSEG_COLD_DATA;
default:
return CURSEG_WARM_DATA;
}
}
/* This returns write hints for each segment type. This hints will be
* passed down to block layer. There are mapping tables which depend on
* the mount option 'whint_mode'.
*
* 1) whint_mode=off. F2FS only passes down WRITE_LIFE_NOT_SET.
*
* 2) whint_mode=user-based. F2FS tries to pass down hints given by users.
*
* User F2FS Block
* ---- ---- -----
* META WRITE_LIFE_NOT_SET
* HOT_NODE "
* WARM_NODE "
* COLD_NODE "
* ioctl(COLD) COLD_DATA WRITE_LIFE_EXTREME
* extension list " "
*
* -- buffered io
* WRITE_LIFE_EXTREME COLD_DATA WRITE_LIFE_EXTREME
* WRITE_LIFE_SHORT HOT_DATA WRITE_LIFE_SHORT
* WRITE_LIFE_NOT_SET WARM_DATA WRITE_LIFE_NOT_SET
* WRITE_LIFE_NONE " "
* WRITE_LIFE_MEDIUM " "
* WRITE_LIFE_LONG " "
*
* -- direct io
* WRITE_LIFE_EXTREME COLD_DATA WRITE_LIFE_EXTREME
* WRITE_LIFE_SHORT HOT_DATA WRITE_LIFE_SHORT
* WRITE_LIFE_NOT_SET WARM_DATA WRITE_LIFE_NOT_SET
* WRITE_LIFE_NONE " WRITE_LIFE_NONE
* WRITE_LIFE_MEDIUM " WRITE_LIFE_MEDIUM
* WRITE_LIFE_LONG " WRITE_LIFE_LONG
*
* 3) whint_mode=fs-based. F2FS passes down hints with its policy.
*
* User F2FS Block
* ---- ---- -----
* META WRITE_LIFE_MEDIUM;
* HOT_NODE WRITE_LIFE_NOT_SET
* WARM_NODE "
* COLD_NODE WRITE_LIFE_NONE
* ioctl(COLD) COLD_DATA WRITE_LIFE_EXTREME
* extension list " "
*
* -- buffered io
* WRITE_LIFE_EXTREME COLD_DATA WRITE_LIFE_EXTREME
* WRITE_LIFE_SHORT HOT_DATA WRITE_LIFE_SHORT
* WRITE_LIFE_NOT_SET WARM_DATA WRITE_LIFE_LONG
* WRITE_LIFE_NONE " "
* WRITE_LIFE_MEDIUM " "
* WRITE_LIFE_LONG " "
*
* -- direct io
* WRITE_LIFE_EXTREME COLD_DATA WRITE_LIFE_EXTREME
* WRITE_LIFE_SHORT HOT_DATA WRITE_LIFE_SHORT
* WRITE_LIFE_NOT_SET WARM_DATA WRITE_LIFE_NOT_SET
* WRITE_LIFE_NONE " WRITE_LIFE_NONE
* WRITE_LIFE_MEDIUM " WRITE_LIFE_MEDIUM
* WRITE_LIFE_LONG " WRITE_LIFE_LONG
*/
enum rw_hint io_type_to_rw_hint(struct f2fs_sb_info *sbi,
enum page_type type, enum temp_type temp)
{
if (F2FS_OPTION(sbi).whint_mode == WHINT_MODE_USER) {
if (type == DATA) {
if (temp == WARM)
return WRITE_LIFE_NOT_SET;
else if (temp == HOT)
return WRITE_LIFE_SHORT;
else if (temp == COLD)
return WRITE_LIFE_EXTREME;
} else {
return WRITE_LIFE_NOT_SET;
}
} else if (F2FS_OPTION(sbi).whint_mode == WHINT_MODE_FS) {
if (type == DATA) {
if (temp == WARM)
return WRITE_LIFE_LONG;
else if (temp == HOT)
return WRITE_LIFE_SHORT;
else if (temp == COLD)
return WRITE_LIFE_EXTREME;
} else if (type == NODE) {
if (temp == WARM || temp == HOT)
return WRITE_LIFE_NOT_SET;
else if (temp == COLD)
return WRITE_LIFE_NONE;
} else if (type == META) {
return WRITE_LIFE_MEDIUM;
}
}
return WRITE_LIFE_NOT_SET;
}
static int __get_segment_type_2(struct f2fs_io_info *fio)
{
if (fio->type == DATA)
return CURSEG_HOT_DATA;
else
return CURSEG_HOT_NODE;
}
static int __get_segment_type_4(struct f2fs_io_info *fio)
{
if (fio->type == DATA) {
struct inode *inode = fio->page->mapping->host;
if (S_ISDIR(inode->i_mode))
return CURSEG_HOT_DATA;
else
return CURSEG_COLD_DATA;
} else {
if (IS_DNODE(fio->page) && is_cold_node(fio->page))
return CURSEG_WARM_NODE;
else
return CURSEG_COLD_NODE;
}
}
static int __get_segment_type_6(struct f2fs_io_info *fio)
{
if (fio->type == DATA) {
struct inode *inode = fio->page->mapping->host;
if (is_cold_data(fio->page) || file_is_cold(inode))
return CURSEG_COLD_DATA;
if (file_is_hot(inode) ||
is_inode_flag_set(inode, FI_HOT_DATA))
return CURSEG_HOT_DATA;
return rw_hint_to_seg_type(inode->i_write_hint);
} else {
if (IS_DNODE(fio->page))
return is_cold_node(fio->page) ? CURSEG_WARM_NODE :
CURSEG_HOT_NODE;
return CURSEG_COLD_NODE;
}
}
static int __get_segment_type(struct f2fs_io_info *fio)
{
int type = 0;
switch (F2FS_OPTION(fio->sbi).active_logs) {
case 2:
type = __get_segment_type_2(fio);
break;
case 4:
type = __get_segment_type_4(fio);
break;
case 6:
type = __get_segment_type_6(fio);
break;
default:
f2fs_bug_on(fio->sbi, true);
}
if (IS_HOT(type))
fio->temp = HOT;
else if (IS_WARM(type))
fio->temp = WARM;
else
fio->temp = COLD;
return type;
}
void allocate_data_block(struct f2fs_sb_info *sbi, struct page *page,
block_t old_blkaddr, block_t *new_blkaddr,
struct f2fs_summary *sum, int type,
struct f2fs_io_info *fio, bool add_list)
{
struct sit_info *sit_i = SIT_I(sbi);
struct curseg_info *curseg = CURSEG_I(sbi, type);
down_read(&SM_I(sbi)->curseg_lock);
mutex_lock(&curseg->curseg_mutex);
down_write(&sit_i->sentry_lock);
*new_blkaddr = NEXT_FREE_BLKADDR(sbi, curseg);
f2fs_wait_discard_bio(sbi, *new_blkaddr);
/*
* __add_sum_entry should be resided under the curseg_mutex
* because, this function updates a summary entry in the
* current summary block.
*/
__add_sum_entry(sbi, type, sum);
__refresh_next_blkoff(sbi, curseg);
stat_inc_block_count(sbi, curseg);
/*
* SIT information should be updated before segment allocation,
* since SSR needs latest valid block information.
*/
update_sit_entry(sbi, *new_blkaddr, 1);
if (GET_SEGNO(sbi, old_blkaddr) != NULL_SEGNO)
update_sit_entry(sbi, old_blkaddr, -1);
if (!__has_curseg_space(sbi, type))
sit_i->s_ops->allocate_segment(sbi, type, false);
/*
* segment dirty status should be updated after segment allocation,
* so we just need to update status only one time after previous
* segment being closed.
*/
locate_dirty_segment(sbi, GET_SEGNO(sbi, old_blkaddr));
locate_dirty_segment(sbi, GET_SEGNO(sbi, *new_blkaddr));
up_write(&sit_i->sentry_lock);
if (page && IS_NODESEG(type)) {
fill_node_footer_blkaddr(page, NEXT_FREE_BLKADDR(sbi, curseg));
f2fs_inode_chksum_set(sbi, page);
}
if (add_list) {
struct f2fs_bio_info *io;
INIT_LIST_HEAD(&fio->list);
fio->in_list = true;
io = sbi->write_io[fio->type] + fio->temp;
spin_lock(&io->io_lock);
list_add_tail(&fio->list, &io->io_list);
spin_unlock(&io->io_lock);
}
mutex_unlock(&curseg->curseg_mutex);
up_read(&SM_I(sbi)->curseg_lock);
}
static void update_device_state(struct f2fs_io_info *fio)
{
struct f2fs_sb_info *sbi = fio->sbi;
unsigned int devidx;
if (!sbi->s_ndevs)
return;
devidx = f2fs_target_device_index(sbi, fio->new_blkaddr);
/* update device state for fsync */
set_dirty_device(sbi, fio->ino, devidx, FLUSH_INO);
/* update device state for checkpoint */
if (!f2fs_test_bit(devidx, (char *)&sbi->dirty_device)) {
spin_lock(&sbi->dev_lock);
f2fs_set_bit(devidx, (char *)&sbi->dirty_device);
spin_unlock(&sbi->dev_lock);
}
}
static void do_write_page(struct f2fs_summary *sum, struct f2fs_io_info *fio)
{
int type = __get_segment_type(fio);
int err;
reallocate:
allocate_data_block(fio->sbi, fio->page, fio->old_blkaddr,
&fio->new_blkaddr, sum, type, fio, true);
/* writeout dirty page into bdev */
err = f2fs_submit_page_write(fio);
if (err == -EAGAIN) {
fio->old_blkaddr = fio->new_blkaddr;
goto reallocate;
} else if (!err) {
update_device_state(fio);
}
}
void write_meta_page(struct f2fs_sb_info *sbi, struct page *page,
enum iostat_type io_type)
{
struct f2fs_io_info fio = {
.sbi = sbi,
.type = META,
.temp = HOT,
.op = REQ_OP_WRITE,
.op_flags = REQ_SYNC | REQ_META | REQ_PRIO,
.old_blkaddr = page->index,
.new_blkaddr = page->index,
.page = page,
.encrypted_page = NULL,
.in_list = false,
};
if (unlikely(page->index >= MAIN_BLKADDR(sbi)))
fio.op_flags &= ~REQ_META;
set_page_writeback(page);
f2fs_submit_page_write(&fio);
f2fs_update_iostat(sbi, io_type, F2FS_BLKSIZE);
}
void write_node_page(unsigned int nid, struct f2fs_io_info *fio)
{
struct f2fs_summary sum;
set_summary(&sum, nid, 0, 0);
do_write_page(&sum, fio);
f2fs_update_iostat(fio->sbi, fio->io_type, F2FS_BLKSIZE);
}
void write_data_page(struct dnode_of_data *dn, struct f2fs_io_info *fio)
{
struct f2fs_sb_info *sbi = fio->sbi;
struct f2fs_summary sum;
struct node_info ni;
f2fs_bug_on(sbi, dn->data_blkaddr == NULL_ADDR);
get_node_info(sbi, dn->nid, &ni);
set_summary(&sum, dn->nid, dn->ofs_in_node, ni.version);
do_write_page(&sum, fio);
f2fs_update_data_blkaddr(dn, fio->new_blkaddr);
f2fs_update_iostat(sbi, fio->io_type, F2FS_BLKSIZE);
}
int rewrite_data_page(struct f2fs_io_info *fio)
{
int err;
struct f2fs_sb_info *sbi = fio->sbi;
fio->new_blkaddr = fio->old_blkaddr;
/* i/o temperature is needed for passing down write hints */
__get_segment_type(fio);
f2fs_bug_on(sbi, !IS_DATASEG(get_seg_entry(sbi,
GET_SEGNO(sbi, fio->new_blkaddr))->type));
stat_inc_inplace_blocks(fio->sbi);
err = f2fs_submit_page_bio(fio);
if (!err)
update_device_state(fio);
f2fs_update_iostat(fio->sbi, fio->io_type, F2FS_BLKSIZE);
return err;
}
static inline int __f2fs_get_curseg(struct f2fs_sb_info *sbi,
unsigned int segno)
{
int i;
for (i = CURSEG_HOT_DATA; i < NO_CHECK_TYPE; i++) {
if (CURSEG_I(sbi, i)->segno == segno)
break;
}
return i;
}
void __f2fs_replace_block(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
block_t old_blkaddr, block_t new_blkaddr,
bool recover_curseg, bool recover_newaddr)
{
struct sit_info *sit_i = SIT_I(sbi);
struct curseg_info *curseg;
unsigned int segno, old_cursegno;
struct seg_entry *se;
int type;
unsigned short old_blkoff;
segno = GET_SEGNO(sbi, new_blkaddr);
se = get_seg_entry(sbi, segno);
type = se->type;
down_write(&SM_I(sbi)->curseg_lock);
if (!recover_curseg) {
/* for recovery flow */
if (se->valid_blocks == 0 && !IS_CURSEG(sbi, segno)) {
if (old_blkaddr == NULL_ADDR)
type = CURSEG_COLD_DATA;
else
type = CURSEG_WARM_DATA;
}
} else {
if (IS_CURSEG(sbi, segno)) {
/* se->type is volatile as SSR allocation */
type = __f2fs_get_curseg(sbi, segno);
f2fs_bug_on(sbi, type == NO_CHECK_TYPE);
} else {
type = CURSEG_WARM_DATA;
}
}
f2fs_bug_on(sbi, !IS_DATASEG(type));
curseg = CURSEG_I(sbi, type);
mutex_lock(&curseg->curseg_mutex);
down_write(&sit_i->sentry_lock);
old_cursegno = curseg->segno;
old_blkoff = curseg->next_blkoff;
/* change the current segment */
if (segno != curseg->segno) {
curseg->next_segno = segno;
change_curseg(sbi, type);
}
curseg->next_blkoff = GET_BLKOFF_FROM_SEG0(sbi, new_blkaddr);
__add_sum_entry(sbi, type, sum);
if (!recover_curseg || recover_newaddr)
update_sit_entry(sbi, new_blkaddr, 1);
if (GET_SEGNO(sbi, old_blkaddr) != NULL_SEGNO)
update_sit_entry(sbi, old_blkaddr, -1);
locate_dirty_segment(sbi, GET_SEGNO(sbi, old_blkaddr));
locate_dirty_segment(sbi, GET_SEGNO(sbi, new_blkaddr));
locate_dirty_segment(sbi, old_cursegno);
if (recover_curseg) {
if (old_cursegno != curseg->segno) {
curseg->next_segno = old_cursegno;
change_curseg(sbi, type);
}
curseg->next_blkoff = old_blkoff;
}
up_write(&sit_i->sentry_lock);
mutex_unlock(&curseg->curseg_mutex);
up_write(&SM_I(sbi)->curseg_lock);
}
void f2fs_replace_block(struct f2fs_sb_info *sbi, struct dnode_of_data *dn,
block_t old_addr, block_t new_addr,
unsigned char version, bool recover_curseg,
bool recover_newaddr)
{
struct f2fs_summary sum;
set_summary(&sum, dn->nid, dn->ofs_in_node, version);
__f2fs_replace_block(sbi, &sum, old_addr, new_addr,
recover_curseg, recover_newaddr);
f2fs_update_data_blkaddr(dn, new_addr);
}
void f2fs_wait_on_page_writeback(struct page *page,
enum page_type type, bool ordered)
{
if (PageWriteback(page)) {
struct f2fs_sb_info *sbi = F2FS_P_SB(page);
f2fs_submit_merged_write_cond(sbi, page->mapping->host,
0, page->index, type);
if (ordered)
wait_on_page_writeback(page);
else
wait_for_stable_page(page);
}
}
void f2fs_wait_on_block_writeback(struct f2fs_sb_info *sbi, block_t blkaddr)
f2fs crypto: fix racing of accessing encrypted page among different competitors Since we use different page cache (normally inode's page cache for R/W and meta inode's page cache for GC) to cache the same physical block which is belong to an encrypted inode. Writeback of these two page cache should be exclusive, but now we didn't handle writeback state well, so there may be potential racing problem: a) kworker: f2fs_gc: - f2fs_write_data_pages - f2fs_write_data_page - do_write_data_page - write_data_page - f2fs_submit_page_mbio (page#1 in inode's page cache was queued in f2fs bio cache, and be ready to write to new blkaddr) - gc_data_segment - move_encrypted_block - pagecache_get_page (page#2 in meta inode's page cache was cached with the invalid datas of physical block located in new blkaddr) - f2fs_submit_page_mbio (page#1 was submitted, later, page#2 with invalid data will be submitted) b) f2fs_gc: - gc_data_segment - move_encrypted_block - f2fs_submit_page_mbio (page#1 in meta inode's page cache was queued in f2fs bio cache, and be ready to write to new blkaddr) user thread: - f2fs_write_begin - f2fs_submit_page_bio (we submit the request to block layer to update page#2 in inode's page cache with physical block located in new blkaddr, so here we may read gabbage data from new blkaddr since GC hasn't writebacked the page#1 yet) This patch fixes above potential racing problem for encrypted inode. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2015-10-08 12:27:34 +07:00
{
struct page *cpage;
if (blkaddr == NEW_ADDR || blkaddr == NULL_ADDR)
f2fs crypto: fix racing of accessing encrypted page among different competitors Since we use different page cache (normally inode's page cache for R/W and meta inode's page cache for GC) to cache the same physical block which is belong to an encrypted inode. Writeback of these two page cache should be exclusive, but now we didn't handle writeback state well, so there may be potential racing problem: a) kworker: f2fs_gc: - f2fs_write_data_pages - f2fs_write_data_page - do_write_data_page - write_data_page - f2fs_submit_page_mbio (page#1 in inode's page cache was queued in f2fs bio cache, and be ready to write to new blkaddr) - gc_data_segment - move_encrypted_block - pagecache_get_page (page#2 in meta inode's page cache was cached with the invalid datas of physical block located in new blkaddr) - f2fs_submit_page_mbio (page#1 was submitted, later, page#2 with invalid data will be submitted) b) f2fs_gc: - gc_data_segment - move_encrypted_block - f2fs_submit_page_mbio (page#1 in meta inode's page cache was queued in f2fs bio cache, and be ready to write to new blkaddr) user thread: - f2fs_write_begin - f2fs_submit_page_bio (we submit the request to block layer to update page#2 in inode's page cache with physical block located in new blkaddr, so here we may read gabbage data from new blkaddr since GC hasn't writebacked the page#1 yet) This patch fixes above potential racing problem for encrypted inode. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2015-10-08 12:27:34 +07:00
return;
cpage = find_lock_page(META_MAPPING(sbi), blkaddr);
if (cpage) {
f2fs_wait_on_page_writeback(cpage, DATA, true);
f2fs crypto: fix racing of accessing encrypted page among different competitors Since we use different page cache (normally inode's page cache for R/W and meta inode's page cache for GC) to cache the same physical block which is belong to an encrypted inode. Writeback of these two page cache should be exclusive, but now we didn't handle writeback state well, so there may be potential racing problem: a) kworker: f2fs_gc: - f2fs_write_data_pages - f2fs_write_data_page - do_write_data_page - write_data_page - f2fs_submit_page_mbio (page#1 in inode's page cache was queued in f2fs bio cache, and be ready to write to new blkaddr) - gc_data_segment - move_encrypted_block - pagecache_get_page (page#2 in meta inode's page cache was cached with the invalid datas of physical block located in new blkaddr) - f2fs_submit_page_mbio (page#1 was submitted, later, page#2 with invalid data will be submitted) b) f2fs_gc: - gc_data_segment - move_encrypted_block - f2fs_submit_page_mbio (page#1 in meta inode's page cache was queued in f2fs bio cache, and be ready to write to new blkaddr) user thread: - f2fs_write_begin - f2fs_submit_page_bio (we submit the request to block layer to update page#2 in inode's page cache with physical block located in new blkaddr, so here we may read gabbage data from new blkaddr since GC hasn't writebacked the page#1 yet) This patch fixes above potential racing problem for encrypted inode. Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2015-10-08 12:27:34 +07:00
f2fs_put_page(cpage, 1);
}
}
static void read_compacted_summaries(struct f2fs_sb_info *sbi)
{
struct f2fs_checkpoint *ckpt = F2FS_CKPT(sbi);
struct curseg_info *seg_i;
unsigned char *kaddr;
struct page *page;
block_t start;
int i, j, offset;
start = start_sum_block(sbi);
page = get_meta_page(sbi, start++);
kaddr = (unsigned char *)page_address(page);
/* Step 1: restore nat cache */
seg_i = CURSEG_I(sbi, CURSEG_HOT_DATA);
memcpy(seg_i->journal, kaddr, SUM_JOURNAL_SIZE);
/* Step 2: restore sit cache */
seg_i = CURSEG_I(sbi, CURSEG_COLD_DATA);
memcpy(seg_i->journal, kaddr + SUM_JOURNAL_SIZE, SUM_JOURNAL_SIZE);
offset = 2 * SUM_JOURNAL_SIZE;
/* Step 3: restore summary entries */
for (i = CURSEG_HOT_DATA; i <= CURSEG_COLD_DATA; i++) {
unsigned short blk_off;
unsigned int segno;
seg_i = CURSEG_I(sbi, i);
segno = le32_to_cpu(ckpt->cur_data_segno[i]);
blk_off = le16_to_cpu(ckpt->cur_data_blkoff[i]);
seg_i->next_segno = segno;
reset_curseg(sbi, i, 0);
seg_i->alloc_type = ckpt->alloc_type[i];
seg_i->next_blkoff = blk_off;
if (seg_i->alloc_type == SSR)
blk_off = sbi->blocks_per_seg;
for (j = 0; j < blk_off; j++) {
struct f2fs_summary *s;
s = (struct f2fs_summary *)(kaddr + offset);
seg_i->sum_blk->entries[j] = *s;
offset += SUMMARY_SIZE;
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 19:29:47 +07:00
if (offset + SUMMARY_SIZE <= PAGE_SIZE -
SUM_FOOTER_SIZE)
continue;
f2fs_put_page(page, 1);
page = NULL;
page = get_meta_page(sbi, start++);
kaddr = (unsigned char *)page_address(page);
offset = 0;
}
}
f2fs_put_page(page, 1);
}
static int read_normal_summaries(struct f2fs_sb_info *sbi, int type)
{
struct f2fs_checkpoint *ckpt = F2FS_CKPT(sbi);
struct f2fs_summary_block *sum;
struct curseg_info *curseg;
struct page *new;
unsigned short blk_off;
unsigned int segno = 0;
block_t blk_addr = 0;
/* get segment number and block addr */
if (IS_DATASEG(type)) {
segno = le32_to_cpu(ckpt->cur_data_segno[type]);
blk_off = le16_to_cpu(ckpt->cur_data_blkoff[type -
CURSEG_HOT_DATA]);
if (__exist_node_summaries(sbi))
blk_addr = sum_blk_addr(sbi, NR_CURSEG_TYPE, type);
else
blk_addr = sum_blk_addr(sbi, NR_CURSEG_DATA_TYPE, type);
} else {
segno = le32_to_cpu(ckpt->cur_node_segno[type -
CURSEG_HOT_NODE]);
blk_off = le16_to_cpu(ckpt->cur_node_blkoff[type -
CURSEG_HOT_NODE]);
if (__exist_node_summaries(sbi))
blk_addr = sum_blk_addr(sbi, NR_CURSEG_NODE_TYPE,
type - CURSEG_HOT_NODE);
else
blk_addr = GET_SUM_BLOCK(sbi, segno);
}
new = get_meta_page(sbi, blk_addr);
sum = (struct f2fs_summary_block *)page_address(new);
if (IS_NODESEG(type)) {
if (__exist_node_summaries(sbi)) {
struct f2fs_summary *ns = &sum->entries[0];
int i;
for (i = 0; i < sbi->blocks_per_seg; i++, ns++) {
ns->version = 0;
ns->ofs_in_node = 0;
}
} else {
restore_node_summary(sbi, segno, sum);
}
}
/* set uncompleted segment to curseg */
curseg = CURSEG_I(sbi, type);
mutex_lock(&curseg->curseg_mutex);
/* update journal info */
down_write(&curseg->journal_rwsem);
memcpy(curseg->journal, &sum->journal, SUM_JOURNAL_SIZE);
up_write(&curseg->journal_rwsem);
memcpy(curseg->sum_blk->entries, sum->entries, SUM_ENTRY_SIZE);
memcpy(&curseg->sum_blk->footer, &sum->footer, SUM_FOOTER_SIZE);
curseg->next_segno = segno;
reset_curseg(sbi, type, 0);
curseg->alloc_type = ckpt->alloc_type[type];
curseg->next_blkoff = blk_off;
mutex_unlock(&curseg->curseg_mutex);
f2fs_put_page(new, 1);
return 0;
}
static int restore_curseg_summaries(struct f2fs_sb_info *sbi)
{
struct f2fs_journal *sit_j = CURSEG_I(sbi, CURSEG_COLD_DATA)->journal;
struct f2fs_journal *nat_j = CURSEG_I(sbi, CURSEG_HOT_DATA)->journal;
int type = CURSEG_HOT_DATA;
int err;
if (is_set_ckpt_flags(sbi, CP_COMPACT_SUM_FLAG)) {
int npages = npages_for_summary_flush(sbi, true);
if (npages >= 2)
ra_meta_pages(sbi, start_sum_block(sbi), npages,
META_CP, true);
/* restore for compacted data summary */
read_compacted_summaries(sbi);
type = CURSEG_HOT_NODE;
}
if (__exist_node_summaries(sbi))
ra_meta_pages(sbi, sum_blk_addr(sbi, NR_CURSEG_TYPE, type),
NR_CURSEG_TYPE - type, META_CP, true);
for (; type <= CURSEG_COLD_NODE; type++) {
err = read_normal_summaries(sbi, type);
if (err)
return err;
}
/* sanity check for summary blocks */
if (nats_in_cursum(nat_j) > NAT_JOURNAL_ENTRIES ||
sits_in_cursum(sit_j) > SIT_JOURNAL_ENTRIES)
return -EINVAL;
return 0;
}
static void write_compacted_summaries(struct f2fs_sb_info *sbi, block_t blkaddr)
{
struct page *page;
unsigned char *kaddr;
struct f2fs_summary *summary;
struct curseg_info *seg_i;
int written_size = 0;
int i, j;
page = grab_meta_page(sbi, blkaddr++);
kaddr = (unsigned char *)page_address(page);
/* Step 1: write nat cache */
seg_i = CURSEG_I(sbi, CURSEG_HOT_DATA);
memcpy(kaddr, seg_i->journal, SUM_JOURNAL_SIZE);
written_size += SUM_JOURNAL_SIZE;
/* Step 2: write sit cache */
seg_i = CURSEG_I(sbi, CURSEG_COLD_DATA);
memcpy(kaddr + written_size, seg_i->journal, SUM_JOURNAL_SIZE);
written_size += SUM_JOURNAL_SIZE;
/* Step 3: write summary entries */
for (i = CURSEG_HOT_DATA; i <= CURSEG_COLD_DATA; i++) {
unsigned short blkoff;
seg_i = CURSEG_I(sbi, i);
if (sbi->ckpt->alloc_type[i] == SSR)
blkoff = sbi->blocks_per_seg;
else
blkoff = curseg_blkoff(sbi, i);
for (j = 0; j < blkoff; j++) {
if (!page) {
page = grab_meta_page(sbi, blkaddr++);
kaddr = (unsigned char *)page_address(page);
written_size = 0;
}
summary = (struct f2fs_summary *)(kaddr + written_size);
*summary = seg_i->sum_blk->entries[j];
written_size += SUMMARY_SIZE;
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 19:29:47 +07:00
if (written_size + SUMMARY_SIZE <= PAGE_SIZE -
SUM_FOOTER_SIZE)
continue;
set_page_dirty(page);
f2fs_put_page(page, 1);
page = NULL;
}
}
if (page) {
set_page_dirty(page);
f2fs_put_page(page, 1);
}
}
static void write_normal_summaries(struct f2fs_sb_info *sbi,
block_t blkaddr, int type)
{
int i, end;
if (IS_DATASEG(type))
end = type + NR_CURSEG_DATA_TYPE;
else
end = type + NR_CURSEG_NODE_TYPE;
for (i = type; i < end; i++)
write_current_sum_page(sbi, i, blkaddr + (i - type));
}
void write_data_summaries(struct f2fs_sb_info *sbi, block_t start_blk)
{
if (is_set_ckpt_flags(sbi, CP_COMPACT_SUM_FLAG))
write_compacted_summaries(sbi, start_blk);
else
write_normal_summaries(sbi, start_blk, CURSEG_HOT_DATA);
}
void write_node_summaries(struct f2fs_sb_info *sbi, block_t start_blk)
{
write_normal_summaries(sbi, start_blk, CURSEG_HOT_NODE);
}
int lookup_journal_in_cursum(struct f2fs_journal *journal, int type,
unsigned int val, int alloc)
{
int i;
if (type == NAT_JOURNAL) {
for (i = 0; i < nats_in_cursum(journal); i++) {
if (le32_to_cpu(nid_in_journal(journal, i)) == val)
return i;
}
if (alloc && __has_cursum_space(journal, 1, NAT_JOURNAL))
return update_nats_in_cursum(journal, 1);
} else if (type == SIT_JOURNAL) {
for (i = 0; i < sits_in_cursum(journal); i++)
if (le32_to_cpu(segno_in_journal(journal, i)) == val)
return i;
if (alloc && __has_cursum_space(journal, 1, SIT_JOURNAL))
return update_sits_in_cursum(journal, 1);
}
return -1;
}
static struct page *get_current_sit_page(struct f2fs_sb_info *sbi,
unsigned int segno)
{
return get_meta_page(sbi, current_sit_addr(sbi, segno));
}
static struct page *get_next_sit_page(struct f2fs_sb_info *sbi,
unsigned int start)
{
struct sit_info *sit_i = SIT_I(sbi);
f2fs: rebuild sit page from sit info in mem This patch rebuild sit page from sit info in mem instead of issue a read io. I test this method and the result is as below: Pre: mmc_perf_test-12061 [001] ...1 976.819992: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [001] ...1 976.856446: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 998.976946: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 999.023269: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 1022.060772: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 1022.111034: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [002] ...1 1070.127643: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 1070.187352: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 1095.942124: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 1095.995975: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 1122.535091: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 1122.586521: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [001] ...1 1147.897487: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [001] ...1 1147.959438: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 1177.926951: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [002] ...1 1177.976823: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [002] ...1 1204.176087: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [002] ...1 1204.239046: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit Some sit flush consume more than 50ms. Now: mmc_perf_test-2187 [007] ...1 196.840684: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [007] ...1 196.841258: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [007] ...1 219.430582: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [007] ...1 219.431144: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [002] ...1 243.638678: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [000] ...1 243.638980: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [002] ...1 265.392180: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [002] ...1 265.392245: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [000] ...1 290.309051: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [000] ...1 290.309116: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [003] ...1 317.144209: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [003] ...1 317.145913: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [005] ...1 343.224954: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [005] ...1 343.225574: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [000] ...1 370.239846: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [000] ...1 370.241138: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [001] ...1 397.029043: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [001] ...1 397.030750: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [003] ...1 425.386377: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [003] ...1 425.387735: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit Most sit flush consume no more than 1ms. Signed-off-by: Yunlei He <heyunlei@huawei.com> Reviewed-by: Chao Yu <yuchao0@huawei.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2018-01-25 16:27:11 +07:00
struct page *page;
pgoff_t src_off, dst_off;
src_off = current_sit_addr(sbi, start);
dst_off = next_sit_addr(sbi, src_off);
f2fs: rebuild sit page from sit info in mem This patch rebuild sit page from sit info in mem instead of issue a read io. I test this method and the result is as below: Pre: mmc_perf_test-12061 [001] ...1 976.819992: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [001] ...1 976.856446: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 998.976946: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 999.023269: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 1022.060772: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 1022.111034: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [002] ...1 1070.127643: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 1070.187352: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 1095.942124: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 1095.995975: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 1122.535091: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 1122.586521: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [001] ...1 1147.897487: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [001] ...1 1147.959438: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 1177.926951: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [002] ...1 1177.976823: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [002] ...1 1204.176087: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [002] ...1 1204.239046: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit Some sit flush consume more than 50ms. Now: mmc_perf_test-2187 [007] ...1 196.840684: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [007] ...1 196.841258: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [007] ...1 219.430582: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [007] ...1 219.431144: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [002] ...1 243.638678: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [000] ...1 243.638980: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [002] ...1 265.392180: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [002] ...1 265.392245: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [000] ...1 290.309051: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [000] ...1 290.309116: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [003] ...1 317.144209: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [003] ...1 317.145913: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [005] ...1 343.224954: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [005] ...1 343.225574: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [000] ...1 370.239846: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [000] ...1 370.241138: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [001] ...1 397.029043: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [001] ...1 397.030750: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [003] ...1 425.386377: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [003] ...1 425.387735: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit Most sit flush consume no more than 1ms. Signed-off-by: Yunlei He <heyunlei@huawei.com> Reviewed-by: Chao Yu <yuchao0@huawei.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2018-01-25 16:27:11 +07:00
page = grab_meta_page(sbi, dst_off);
seg_info_to_sit_page(sbi, page, start);
f2fs: rebuild sit page from sit info in mem This patch rebuild sit page from sit info in mem instead of issue a read io. I test this method and the result is as below: Pre: mmc_perf_test-12061 [001] ...1 976.819992: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [001] ...1 976.856446: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 998.976946: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 999.023269: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 1022.060772: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 1022.111034: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [002] ...1 1070.127643: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 1070.187352: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 1095.942124: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 1095.995975: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 1122.535091: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 1122.586521: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [001] ...1 1147.897487: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [001] ...1 1147.959438: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 1177.926951: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [002] ...1 1177.976823: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [002] ...1 1204.176087: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [002] ...1 1204.239046: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit Some sit flush consume more than 50ms. Now: mmc_perf_test-2187 [007] ...1 196.840684: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [007] ...1 196.841258: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [007] ...1 219.430582: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [007] ...1 219.431144: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [002] ...1 243.638678: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [000] ...1 243.638980: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [002] ...1 265.392180: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [002] ...1 265.392245: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [000] ...1 290.309051: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [000] ...1 290.309116: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [003] ...1 317.144209: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [003] ...1 317.145913: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [005] ...1 343.224954: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [005] ...1 343.225574: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [000] ...1 370.239846: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [000] ...1 370.241138: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [001] ...1 397.029043: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [001] ...1 397.030750: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [003] ...1 425.386377: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [003] ...1 425.387735: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit Most sit flush consume no more than 1ms. Signed-off-by: Yunlei He <heyunlei@huawei.com> Reviewed-by: Chao Yu <yuchao0@huawei.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2018-01-25 16:27:11 +07:00
set_page_dirty(page);
set_to_next_sit(sit_i, start);
f2fs: rebuild sit page from sit info in mem This patch rebuild sit page from sit info in mem instead of issue a read io. I test this method and the result is as below: Pre: mmc_perf_test-12061 [001] ...1 976.819992: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [001] ...1 976.856446: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 998.976946: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 999.023269: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 1022.060772: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 1022.111034: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [002] ...1 1070.127643: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 1070.187352: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 1095.942124: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 1095.995975: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 1122.535091: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [003] ...1 1122.586521: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [001] ...1 1147.897487: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [001] ...1 1147.959438: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [003] ...1 1177.926951: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [002] ...1 1177.976823: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-12061 [002] ...1 1204.176087: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-12061 [002] ...1 1204.239046: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit Some sit flush consume more than 50ms. Now: mmc_perf_test-2187 [007] ...1 196.840684: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [007] ...1 196.841258: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [007] ...1 219.430582: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [007] ...1 219.431144: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [002] ...1 243.638678: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [000] ...1 243.638980: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [002] ...1 265.392180: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [002] ...1 265.392245: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [000] ...1 290.309051: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [000] ...1 290.309116: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [003] ...1 317.144209: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [003] ...1 317.145913: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [005] ...1 343.224954: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [005] ...1 343.225574: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [000] ...1 370.239846: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [000] ...1 370.241138: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [001] ...1 397.029043: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [001] ...1 397.030750: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit mmc_perf_test-2187 [003] ...1 425.386377: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = start flush sit mmc_perf_test-2187 [003] ...1 425.387735: f2fs_write_checkpoint: dev = (259,44), checkpoint for Sync, state = end flush sit Most sit flush consume no more than 1ms. Signed-off-by: Yunlei He <heyunlei@huawei.com> Reviewed-by: Chao Yu <yuchao0@huawei.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2018-01-25 16:27:11 +07:00
return page;
}
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
static struct sit_entry_set *grab_sit_entry_set(void)
{
struct sit_entry_set *ses =
f2fs_kmem_cache_alloc(sit_entry_set_slab, GFP_NOFS);
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
ses->entry_cnt = 0;
INIT_LIST_HEAD(&ses->set_list);
return ses;
}
static void release_sit_entry_set(struct sit_entry_set *ses)
{
list_del(&ses->set_list);
kmem_cache_free(sit_entry_set_slab, ses);
}
static void adjust_sit_entry_set(struct sit_entry_set *ses,
struct list_head *head)
{
struct sit_entry_set *next = ses;
if (list_is_last(&ses->set_list, head))
return;
list_for_each_entry_continue(next, head, set_list)
if (ses->entry_cnt <= next->entry_cnt)
break;
list_move_tail(&ses->set_list, &next->set_list);
}
static void add_sit_entry(unsigned int segno, struct list_head *head)
{
struct sit_entry_set *ses;
unsigned int start_segno = START_SEGNO(segno);
list_for_each_entry(ses, head, set_list) {
if (ses->start_segno == start_segno) {
ses->entry_cnt++;
adjust_sit_entry_set(ses, head);
return;
}
}
ses = grab_sit_entry_set();
ses->start_segno = start_segno;
ses->entry_cnt++;
list_add(&ses->set_list, head);
}
static void add_sits_in_set(struct f2fs_sb_info *sbi)
{
struct f2fs_sm_info *sm_info = SM_I(sbi);
struct list_head *set_list = &sm_info->sit_entry_set;
unsigned long *bitmap = SIT_I(sbi)->dirty_sentries_bitmap;
unsigned int segno;
for_each_set_bit(segno, bitmap, MAIN_SEGS(sbi))
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
add_sit_entry(segno, set_list);
}
static void remove_sits_in_journal(struct f2fs_sb_info *sbi)
{
struct curseg_info *curseg = CURSEG_I(sbi, CURSEG_COLD_DATA);
struct f2fs_journal *journal = curseg->journal;
int i;
down_write(&curseg->journal_rwsem);
for (i = 0; i < sits_in_cursum(journal); i++) {
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
unsigned int segno;
bool dirtied;
segno = le32_to_cpu(segno_in_journal(journal, i));
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
dirtied = __mark_sit_entry_dirty(sbi, segno);
if (!dirtied)
add_sit_entry(segno, &SM_I(sbi)->sit_entry_set);
}
update_sits_in_cursum(journal, -i);
up_write(&curseg->journal_rwsem);
}
/*
* CP calls this function, which flushes SIT entries including sit_journal,
* and moves prefree segs to free segs.
*/
void flush_sit_entries(struct f2fs_sb_info *sbi, struct cp_control *cpc)
{
struct sit_info *sit_i = SIT_I(sbi);
unsigned long *bitmap = sit_i->dirty_sentries_bitmap;
struct curseg_info *curseg = CURSEG_I(sbi, CURSEG_COLD_DATA);
struct f2fs_journal *journal = curseg->journal;
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
struct sit_entry_set *ses, *tmp;
struct list_head *head = &SM_I(sbi)->sit_entry_set;
bool to_journal = true;
struct seg_entry *se;
down_write(&sit_i->sentry_lock);
if (!sit_i->dirty_sentries)
goto out;
/*
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
* add and account sit entries of dirty bitmap in sit entry
* set temporarily
*/
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
add_sits_in_set(sbi);
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
/*
* if there are no enough space in journal to store dirty sit
* entries, remove all entries from journal and add and account
* them in sit entry set.
*/
if (!__has_cursum_space(journal, sit_i->dirty_sentries, SIT_JOURNAL))
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
remove_sits_in_journal(sbi);
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
/*
* there are two steps to flush sit entries:
* #1, flush sit entries to journal in current cold data summary block.
* #2, flush sit entries to sit page.
*/
list_for_each_entry_safe(ses, tmp, head, set_list) {
struct page *page = NULL;
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
struct f2fs_sit_block *raw_sit = NULL;
unsigned int start_segno = ses->start_segno;
unsigned int end = min(start_segno + SIT_ENTRY_PER_BLOCK,
(unsigned long)MAIN_SEGS(sbi));
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
unsigned int segno = start_segno;
if (to_journal &&
!__has_cursum_space(journal, ses->entry_cnt, SIT_JOURNAL))
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
to_journal = false;
if (to_journal) {
down_write(&curseg->journal_rwsem);
} else {
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
page = get_next_sit_page(sbi, start_segno);
raw_sit = page_address(page);
}
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
/* flush dirty sit entries in region of current sit set */
for_each_set_bit_from(segno, bitmap, end) {
int offset, sit_offset;
se = get_seg_entry(sbi, segno);
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
/* add discard candidates */
if (!(cpc->reason & CP_DISCARD)) {
cpc->trim_start = segno;
add_discard_addrs(sbi, cpc, false);
}
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
if (to_journal) {
offset = lookup_journal_in_cursum(journal,
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
SIT_JOURNAL, segno, 1);
f2fs_bug_on(sbi, offset < 0);
segno_in_journal(journal, offset) =
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
cpu_to_le32(segno);
seg_info_to_raw_sit(se,
&sit_in_journal(journal, offset));
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
} else {
sit_offset = SIT_ENTRY_OFFSET(sit_i, segno);
seg_info_to_raw_sit(se,
&raw_sit->entries[sit_offset]);
}
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
__clear_bit(segno, bitmap);
sit_i->dirty_sentries--;
ses->entry_cnt--;
}
if (to_journal)
up_write(&curseg->journal_rwsem);
else
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
f2fs_put_page(page, 1);
f2fs_bug_on(sbi, ses->entry_cnt);
release_sit_entry_set(ses);
}
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
f2fs_bug_on(sbi, !list_empty(head));
f2fs_bug_on(sbi, sit_i->dirty_sentries);
out:
if (cpc->reason & CP_DISCARD) {
__u64 trim_start = cpc->trim_start;
for (; cpc->trim_start <= cpc->trim_end; cpc->trim_start++)
add_discard_addrs(sbi, cpc, false);
cpc->trim_start = trim_start;
}
up_write(&sit_i->sentry_lock);
set_prefree_as_free_segments(sbi);
}
static int build_sit_info(struct f2fs_sb_info *sbi)
{
struct f2fs_super_block *raw_super = F2FS_RAW_SUPER(sbi);
struct sit_info *sit_i;
unsigned int sit_segs, start;
char *src_bitmap;
unsigned int bitmap_size;
/* allocate memory for SIT information */
sit_i = f2fs_kzalloc(sbi, sizeof(struct sit_info), GFP_KERNEL);
if (!sit_i)
return -ENOMEM;
SM_I(sbi)->sit_info = sit_i;
sit_i->sentries = f2fs_kvzalloc(sbi, MAIN_SEGS(sbi) *
sizeof(struct seg_entry), GFP_KERNEL);
if (!sit_i->sentries)
return -ENOMEM;
bitmap_size = f2fs_bitmap_size(MAIN_SEGS(sbi));
sit_i->dirty_sentries_bitmap = f2fs_kvzalloc(sbi, bitmap_size,
GFP_KERNEL);
if (!sit_i->dirty_sentries_bitmap)
return -ENOMEM;
for (start = 0; start < MAIN_SEGS(sbi); start++) {
sit_i->sentries[start].cur_valid_map
= f2fs_kzalloc(sbi, SIT_VBLOCK_MAP_SIZE, GFP_KERNEL);
sit_i->sentries[start].ckpt_valid_map
= f2fs_kzalloc(sbi, SIT_VBLOCK_MAP_SIZE, GFP_KERNEL);
if (!sit_i->sentries[start].cur_valid_map ||
!sit_i->sentries[start].ckpt_valid_map)
return -ENOMEM;
#ifdef CONFIG_F2FS_CHECK_FS
sit_i->sentries[start].cur_valid_map_mir
= f2fs_kzalloc(sbi, SIT_VBLOCK_MAP_SIZE, GFP_KERNEL);
if (!sit_i->sentries[start].cur_valid_map_mir)
return -ENOMEM;
#endif
if (f2fs_discard_en(sbi)) {
sit_i->sentries[start].discard_map
= f2fs_kzalloc(sbi, SIT_VBLOCK_MAP_SIZE,
GFP_KERNEL);
if (!sit_i->sentries[start].discard_map)
return -ENOMEM;
}
}
sit_i->tmp_map = f2fs_kzalloc(sbi, SIT_VBLOCK_MAP_SIZE, GFP_KERNEL);
if (!sit_i->tmp_map)
return -ENOMEM;
if (sbi->segs_per_sec > 1) {
sit_i->sec_entries = f2fs_kvzalloc(sbi, MAIN_SECS(sbi) *
sizeof(struct sec_entry), GFP_KERNEL);
if (!sit_i->sec_entries)
return -ENOMEM;
}
/* get information related with SIT */
sit_segs = le32_to_cpu(raw_super->segment_count_sit) >> 1;
/* setup SIT bitmap from ckeckpoint pack */
bitmap_size = __bitmap_size(sbi, SIT_BITMAP);
src_bitmap = __bitmap_ptr(sbi, SIT_BITMAP);
sit_i->sit_bitmap = kmemdup(src_bitmap, bitmap_size, GFP_KERNEL);
if (!sit_i->sit_bitmap)
return -ENOMEM;
#ifdef CONFIG_F2FS_CHECK_FS
sit_i->sit_bitmap_mir = kmemdup(src_bitmap, bitmap_size, GFP_KERNEL);
if (!sit_i->sit_bitmap_mir)
return -ENOMEM;
#endif
/* init SIT information */
sit_i->s_ops = &default_salloc_ops;
sit_i->sit_base_addr = le32_to_cpu(raw_super->sit_blkaddr);
sit_i->sit_blocks = sit_segs << sbi->log_blocks_per_seg;
sit_i->written_valid_blocks = 0;
sit_i->bitmap_size = bitmap_size;
sit_i->dirty_sentries = 0;
sit_i->sents_per_block = SIT_ENTRY_PER_BLOCK;
sit_i->elapsed_time = le64_to_cpu(sbi->ckpt->elapsed_time);
sit_i->mounted_time = ktime_get_real_seconds();
init_rwsem(&sit_i->sentry_lock);
return 0;
}
static int build_free_segmap(struct f2fs_sb_info *sbi)
{
struct free_segmap_info *free_i;
unsigned int bitmap_size, sec_bitmap_size;
/* allocate memory for free segmap information */
free_i = f2fs_kzalloc(sbi, sizeof(struct free_segmap_info), GFP_KERNEL);
if (!free_i)
return -ENOMEM;
SM_I(sbi)->free_info = free_i;
bitmap_size = f2fs_bitmap_size(MAIN_SEGS(sbi));
free_i->free_segmap = f2fs_kvmalloc(sbi, bitmap_size, GFP_KERNEL);
if (!free_i->free_segmap)
return -ENOMEM;
sec_bitmap_size = f2fs_bitmap_size(MAIN_SECS(sbi));
free_i->free_secmap = f2fs_kvmalloc(sbi, sec_bitmap_size, GFP_KERNEL);
if (!free_i->free_secmap)
return -ENOMEM;
/* set all segments as dirty temporarily */
memset(free_i->free_segmap, 0xff, bitmap_size);
memset(free_i->free_secmap, 0xff, sec_bitmap_size);
/* init free segmap information */
free_i->start_segno = GET_SEGNO_FROM_SEG0(sbi, MAIN_BLKADDR(sbi));
free_i->free_segments = 0;
free_i->free_sections = 0;
spin_lock_init(&free_i->segmap_lock);
return 0;
}
static int build_curseg(struct f2fs_sb_info *sbi)
{
struct curseg_info *array;
int i;
array = f2fs_kzalloc(sbi, sizeof(*array) * NR_CURSEG_TYPE, GFP_KERNEL);
if (!array)
return -ENOMEM;
SM_I(sbi)->curseg_array = array;
for (i = 0; i < NR_CURSEG_TYPE; i++) {
mutex_init(&array[i].curseg_mutex);
array[i].sum_blk = f2fs_kzalloc(sbi, PAGE_SIZE, GFP_KERNEL);
if (!array[i].sum_blk)
return -ENOMEM;
init_rwsem(&array[i].journal_rwsem);
array[i].journal = f2fs_kzalloc(sbi,
sizeof(struct f2fs_journal), GFP_KERNEL);
if (!array[i].journal)
return -ENOMEM;
array[i].segno = NULL_SEGNO;
array[i].next_blkoff = 0;
}
return restore_curseg_summaries(sbi);
}
static int build_sit_entries(struct f2fs_sb_info *sbi)
{
struct sit_info *sit_i = SIT_I(sbi);
struct curseg_info *curseg = CURSEG_I(sbi, CURSEG_COLD_DATA);
struct f2fs_journal *journal = curseg->journal;
struct seg_entry *se;
struct f2fs_sit_entry sit;
int sit_blk_cnt = SIT_BLK_CNT(sbi);
unsigned int i, start, end;
unsigned int readed, start_blk = 0;
int err = 0;
do {
readed = ra_meta_pages(sbi, start_blk, BIO_MAX_PAGES,
META_SIT, true);
start = start_blk * sit_i->sents_per_block;
end = (start_blk + readed) * sit_i->sents_per_block;
for (; start < end && start < MAIN_SEGS(sbi); start++) {
struct f2fs_sit_block *sit_blk;
struct page *page;
se = &sit_i->sentries[start];
page = get_current_sit_page(sbi, start);
sit_blk = (struct f2fs_sit_block *)page_address(page);
sit = sit_blk->entries[SIT_ENTRY_OFFSET(sit_i, start)];
f2fs_put_page(page, 1);
err = check_block_count(sbi, start, &sit);
if (err)
return err;
seg_info_from_raw_sit(se, &sit);
/* build discard map only one time */
if (f2fs_discard_en(sbi)) {
if (is_set_ckpt_flags(sbi, CP_TRIMMED_FLAG)) {
memset(se->discard_map, 0xff,
SIT_VBLOCK_MAP_SIZE);
} else {
memcpy(se->discard_map,
se->cur_valid_map,
SIT_VBLOCK_MAP_SIZE);
sbi->discard_blks +=
sbi->blocks_per_seg -
se->valid_blocks;
}
}
if (sbi->segs_per_sec > 1)
get_sec_entry(sbi, start)->valid_blocks +=
se->valid_blocks;
}
start_blk += readed;
} while (start_blk < sit_blk_cnt);
down_read(&curseg->journal_rwsem);
for (i = 0; i < sits_in_cursum(journal); i++) {
unsigned int old_valid_blocks;
start = le32_to_cpu(segno_in_journal(journal, i));
se = &sit_i->sentries[start];
sit = sit_in_journal(journal, i);
old_valid_blocks = se->valid_blocks;
err = check_block_count(sbi, start, &sit);
if (err)
break;
seg_info_from_raw_sit(se, &sit);
if (f2fs_discard_en(sbi)) {
if (is_set_ckpt_flags(sbi, CP_TRIMMED_FLAG)) {
memset(se->discard_map, 0xff,
SIT_VBLOCK_MAP_SIZE);
} else {
memcpy(se->discard_map, se->cur_valid_map,
SIT_VBLOCK_MAP_SIZE);
sbi->discard_blks += old_valid_blocks -
se->valid_blocks;
}
}
if (sbi->segs_per_sec > 1)
get_sec_entry(sbi, start)->valid_blocks +=
se->valid_blocks - old_valid_blocks;
}
up_read(&curseg->journal_rwsem);
return err;
}
static void init_free_segmap(struct f2fs_sb_info *sbi)
{
unsigned int start;
int type;
for (start = 0; start < MAIN_SEGS(sbi); start++) {
struct seg_entry *sentry = get_seg_entry(sbi, start);
if (!sentry->valid_blocks)
__set_free(sbi, start);
else
SIT_I(sbi)->written_valid_blocks +=
sentry->valid_blocks;
}
/* set use the current segments */
for (type = CURSEG_HOT_DATA; type <= CURSEG_COLD_NODE; type++) {
struct curseg_info *curseg_t = CURSEG_I(sbi, type);
__set_test_and_inuse(sbi, curseg_t->segno);
}
}
static void init_dirty_segmap(struct f2fs_sb_info *sbi)
{
struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
struct free_segmap_info *free_i = FREE_I(sbi);
unsigned int segno = 0, offset = 0;
unsigned short valid_blocks;
while (1) {
/* find dirty segment based on free segmap */
segno = find_next_inuse(free_i, MAIN_SEGS(sbi), offset);
if (segno >= MAIN_SEGS(sbi))
break;
offset = segno + 1;
valid_blocks = get_valid_blocks(sbi, segno, false);
if (valid_blocks == sbi->blocks_per_seg || !valid_blocks)
continue;
if (valid_blocks > sbi->blocks_per_seg) {
f2fs_bug_on(sbi, 1);
continue;
}
mutex_lock(&dirty_i->seglist_lock);
__locate_dirty_segment(sbi, segno, DIRTY);
mutex_unlock(&dirty_i->seglist_lock);
}
}
static int init_victim_secmap(struct f2fs_sb_info *sbi)
{
struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
unsigned int bitmap_size = f2fs_bitmap_size(MAIN_SECS(sbi));
dirty_i->victim_secmap = f2fs_kvzalloc(sbi, bitmap_size, GFP_KERNEL);
if (!dirty_i->victim_secmap)
return -ENOMEM;
return 0;
}
static int build_dirty_segmap(struct f2fs_sb_info *sbi)
{
struct dirty_seglist_info *dirty_i;
unsigned int bitmap_size, i;
/* allocate memory for dirty segments list information */
dirty_i = f2fs_kzalloc(sbi, sizeof(struct dirty_seglist_info),
GFP_KERNEL);
if (!dirty_i)
return -ENOMEM;
SM_I(sbi)->dirty_info = dirty_i;
mutex_init(&dirty_i->seglist_lock);
bitmap_size = f2fs_bitmap_size(MAIN_SEGS(sbi));
for (i = 0; i < NR_DIRTY_TYPE; i++) {
dirty_i->dirty_segmap[i] = f2fs_kvzalloc(sbi, bitmap_size,
GFP_KERNEL);
if (!dirty_i->dirty_segmap[i])
return -ENOMEM;
}
init_dirty_segmap(sbi);
return init_victim_secmap(sbi);
}
/*
* Update min, max modified time for cost-benefit GC algorithm
*/
static void init_min_max_mtime(struct f2fs_sb_info *sbi)
{
struct sit_info *sit_i = SIT_I(sbi);
unsigned int segno;
down_write(&sit_i->sentry_lock);
sit_i->min_mtime = LLONG_MAX;
for (segno = 0; segno < MAIN_SEGS(sbi); segno += sbi->segs_per_sec) {
unsigned int i;
unsigned long long mtime = 0;
for (i = 0; i < sbi->segs_per_sec; i++)
mtime += get_seg_entry(sbi, segno + i)->mtime;
mtime = div_u64(mtime, sbi->segs_per_sec);
if (sit_i->min_mtime > mtime)
sit_i->min_mtime = mtime;
}
sit_i->max_mtime = get_mtime(sbi);
up_write(&sit_i->sentry_lock);
}
int build_segment_manager(struct f2fs_sb_info *sbi)
{
struct f2fs_super_block *raw_super = F2FS_RAW_SUPER(sbi);
struct f2fs_checkpoint *ckpt = F2FS_CKPT(sbi);
struct f2fs_sm_info *sm_info;
int err;
sm_info = f2fs_kzalloc(sbi, sizeof(struct f2fs_sm_info), GFP_KERNEL);
if (!sm_info)
return -ENOMEM;
/* init sm info */
sbi->sm_info = sm_info;
sm_info->seg0_blkaddr = le32_to_cpu(raw_super->segment0_blkaddr);
sm_info->main_blkaddr = le32_to_cpu(raw_super->main_blkaddr);
sm_info->segment_count = le32_to_cpu(raw_super->segment_count);
sm_info->reserved_segments = le32_to_cpu(ckpt->rsvd_segment_count);
sm_info->ovp_segments = le32_to_cpu(ckpt->overprov_segment_count);
sm_info->main_segments = le32_to_cpu(raw_super->segment_count_main);
sm_info->ssa_blkaddr = le32_to_cpu(raw_super->ssa_blkaddr);
sm_info->rec_prefree_segments = sm_info->main_segments *
DEF_RECLAIM_PREFREE_SEGMENTS / 100;
if (sm_info->rec_prefree_segments > DEF_MAX_RECLAIM_PREFREE_SEGMENTS)
sm_info->rec_prefree_segments = DEF_MAX_RECLAIM_PREFREE_SEGMENTS;
if (!test_opt(sbi, LFS))
sm_info->ipu_policy = 1 << F2FS_IPU_FSYNC;
sm_info->min_ipu_util = DEF_MIN_IPU_UTIL;
sm_info->min_fsync_blocks = DEF_MIN_FSYNC_BLOCKS;
sm_info->min_hot_blocks = DEF_MIN_HOT_BLOCKS;
sm_info->min_ssr_sections = reserved_sections(sbi);
sm_info->trim_sections = DEF_BATCHED_TRIM_SECTIONS;
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
INIT_LIST_HEAD(&sm_info->sit_entry_set);
init_rwsem(&sm_info->curseg_lock);
if (!f2fs_readonly(sbi->sb)) {
err = create_flush_cmd_control(sbi);
if (err)
return err;
}
err = create_discard_cmd_control(sbi);
if (err)
return err;
err = build_sit_info(sbi);
if (err)
return err;
err = build_free_segmap(sbi);
if (err)
return err;
err = build_curseg(sbi);
if (err)
return err;
/* reinit free segmap based on SIT */
err = build_sit_entries(sbi);
if (err)
return err;
init_free_segmap(sbi);
err = build_dirty_segmap(sbi);
if (err)
return err;
init_min_max_mtime(sbi);
return 0;
}
static void discard_dirty_segmap(struct f2fs_sb_info *sbi,
enum dirty_type dirty_type)
{
struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
mutex_lock(&dirty_i->seglist_lock);
kvfree(dirty_i->dirty_segmap[dirty_type]);
dirty_i->nr_dirty[dirty_type] = 0;
mutex_unlock(&dirty_i->seglist_lock);
}
static void destroy_victim_secmap(struct f2fs_sb_info *sbi)
{
struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
kvfree(dirty_i->victim_secmap);
}
static void destroy_dirty_segmap(struct f2fs_sb_info *sbi)
{
struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
int i;
if (!dirty_i)
return;
/* discard pre-free/dirty segments list */
for (i = 0; i < NR_DIRTY_TYPE; i++)
discard_dirty_segmap(sbi, i);
destroy_victim_secmap(sbi);
SM_I(sbi)->dirty_info = NULL;
kfree(dirty_i);
}
static void destroy_curseg(struct f2fs_sb_info *sbi)
{
struct curseg_info *array = SM_I(sbi)->curseg_array;
int i;
if (!array)
return;
SM_I(sbi)->curseg_array = NULL;
for (i = 0; i < NR_CURSEG_TYPE; i++) {
kfree(array[i].sum_blk);
kfree(array[i].journal);
}
kfree(array);
}
static void destroy_free_segmap(struct f2fs_sb_info *sbi)
{
struct free_segmap_info *free_i = SM_I(sbi)->free_info;
if (!free_i)
return;
SM_I(sbi)->free_info = NULL;
kvfree(free_i->free_segmap);
kvfree(free_i->free_secmap);
kfree(free_i);
}
static void destroy_sit_info(struct f2fs_sb_info *sbi)
{
struct sit_info *sit_i = SIT_I(sbi);
unsigned int start;
if (!sit_i)
return;
if (sit_i->sentries) {
for (start = 0; start < MAIN_SEGS(sbi); start++) {
kfree(sit_i->sentries[start].cur_valid_map);
#ifdef CONFIG_F2FS_CHECK_FS
kfree(sit_i->sentries[start].cur_valid_map_mir);
#endif
kfree(sit_i->sentries[start].ckpt_valid_map);
kfree(sit_i->sentries[start].discard_map);
}
}
kfree(sit_i->tmp_map);
kvfree(sit_i->sentries);
kvfree(sit_i->sec_entries);
kvfree(sit_i->dirty_sentries_bitmap);
SM_I(sbi)->sit_info = NULL;
kfree(sit_i->sit_bitmap);
#ifdef CONFIG_F2FS_CHECK_FS
kfree(sit_i->sit_bitmap_mir);
#endif
kfree(sit_i);
}
void destroy_segment_manager(struct f2fs_sb_info *sbi)
{
struct f2fs_sm_info *sm_info = SM_I(sbi);
if (!sm_info)
return;
destroy_flush_cmd_control(sbi, true);
destroy_discard_cmd_control(sbi);
destroy_dirty_segmap(sbi);
destroy_curseg(sbi);
destroy_free_segmap(sbi);
destroy_sit_info(sbi);
sbi->sm_info = NULL;
kfree(sm_info);
}
int __init create_segment_manager_caches(void)
{
discard_entry_slab = f2fs_kmem_cache_create("discard_entry",
sizeof(struct discard_entry));
if (!discard_entry_slab)
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
goto fail;
discard_cmd_slab = f2fs_kmem_cache_create("discard_cmd",
sizeof(struct discard_cmd));
if (!discard_cmd_slab)
goto destroy_discard_entry;
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
sit_entry_set_slab = f2fs_kmem_cache_create("sit_entry_set",
sizeof(struct sit_entry_set));
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
if (!sit_entry_set_slab)
goto destroy_discard_cmd;
inmem_entry_slab = f2fs_kmem_cache_create("inmem_page_entry",
sizeof(struct inmem_pages));
if (!inmem_entry_slab)
goto destroy_sit_entry_set;
return 0;
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
destroy_sit_entry_set:
kmem_cache_destroy(sit_entry_set_slab);
destroy_discard_cmd:
kmem_cache_destroy(discard_cmd_slab);
destroy_discard_entry:
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
kmem_cache_destroy(discard_entry_slab);
fail:
return -ENOMEM;
}
void destroy_segment_manager_caches(void)
{
f2fs: refactor flush_sit_entries codes for reducing SIT writes In commit aec71382c681 ("f2fs: refactor flush_nat_entries codes for reducing NAT writes"), we descripte the issue as below: "Although building NAT journal in cursum reduce the read/write work for NAT block, but previous design leave us lower performance when write checkpoint frequently for these cases: 1. if journal in cursum has already full, it's a bit of waste that we flush all nat entries to page for persistence, but not to cache any entries. 2. if journal in cursum is not full, we fill nat entries to journal util journal is full, then flush the left dirty entries to disk without merge journaled entries, so these journaled entries may be flushed to disk at next checkpoint but lost chance to flushed last time." Actually, we have the same problem in using SIT journal area. In this patch, firstly we will update sit journal with dirty entries as many as possible. Secondly if there is no space in sit journal, we will remove all entries in journal and walk through the whole dirty entry bitmap of sit, accounting dirty sit entries located in same SIT block to sit entry set. All entry sets are linked to list sit_entry_set in sm_info, sorted ascending order by count of entries in set. Later we flush entries in set which have fewest entries into journal as many as we can, and then flush dense set with merged entries to disk. In this way we can use sit journal area more effectively, also we will reduce SIT update, result in gaining in performance and saving lifetime of flash device. In my testing environment, it shows this patch can help to reduce SIT block update obviously. virtual machine + hard disk: fsstress -p 20 -n 400 -l 5 sit page num cp count sit pages/cp based 2006.50 1349.75 1.486 patched 1566.25 1463.25 1.070 Our latency of merging op is small when handling a great number of dirty SIT entries in flush_sit_entries: latency(ns) dirty sit count 36038 2151 49168 2123 37174 2232 Signed-off-by: Chao Yu <chao2.yu@samsung.com> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
2014-09-04 17:13:01 +07:00
kmem_cache_destroy(sit_entry_set_slab);
kmem_cache_destroy(discard_cmd_slab);
kmem_cache_destroy(discard_entry_slab);
kmem_cache_destroy(inmem_entry_slab);
}