f2fs: turn off atomic writes when deteting abnormal behaviors

It turns out losing meta pages in shutdown period makes f2fs very unstable so that I could see many unexpected error conditions.

Let's keep meta pages for fault injection and sudden power-off tests.

Signed-off-by Jaegeuk Kim <[email protected]>

---
 fs/f2fs/checkpoint.c | 28 +++++++++++++---------------
 1 file changed, 13 insertions(+), 15 deletions(-)

fs/f2fs/checkpoint.c +12 -14

index 9f1c96caebda..37ec0f16448e 100644
@@ -242,11 +242,8 @@ static int __f2fs_write_meta_page(struct page *page,
 
 	trace_f2fs_writepage(page, META);
 
-	if (unlikely(f2fs_cp_error(sbi))) {
-		dec_page_count(sbi, F2FS_DIRTY_META);
-		unlock_page(page);
-		return 0;
-	}
+	if (unlikely(f2fs_cp_error(sbi)))
+		goto redirty_out;
 	if (unlikely(is_sbi_flag_set(sbi, SBI_POR_DOING)))
 		goto redirty_out;
 	if (wbc->for_reclaim && page->index < GET_SUM_BLOCK(sbi, 0))

@@ -1129,6 +1126,9 @@ static void wait_on_all_pages_writeback(struct f2fs_sb_info *sbi)
 		if (!get_pages(sbi, F2FS_WB_CP_DATA))
 			break;
 
+		if (unlikely(f2fs_cp_error(sbi)))
+			break;
+
 		io_schedule_timeout(5*HZ);
 	}
 	finish_wait(&sbi->cp_wait, &wait);

@@ -1202,8 +1202,12 @@ static void commit_checkpoint(struct f2fs_sb_info *sbi,
 
 	/* writeout cp pack 2 page */
 	err = __f2fs_write_meta_page(page, &wbc, FS_CP_META_IO);
-	f2fs_bug_on(sbi, err);
+	if (unlikely(err && f2fs_cp_error(sbi))) {
+		f2fs_put_page(page, 1);
+		return;
+	}
 
+	f2fs_bug_on(sbi, err);
 	f2fs_put_page(page, 0);
 
 	/* submit checkpoint (with barrier if NOBARRIER is not set) */

@@ -1229,7 +1233,7 @@ static int do_checkpoint(struct f2fs_sb_info *sbi, struct cp_control *cpc)
 	while (get_pages(sbi, F2FS_DIRTY_META)) {
 		f2fs_sync_meta_pages(sbi, META, LONG_MAX, FS_CP_META_IO);
 		if (unlikely(f2fs_cp_error(sbi)))
-			return -EIO;
+			break;
 	}
 
 	/*

@@ -1309,7 +1313,7 @@ static int do_checkpoint(struct f2fs_sb_info *sbi, struct cp_control *cpc)
 			f2fs_sync_meta_pages(sbi, META, LONG_MAX,
 							FS_CP_META_IO);
 			if (unlikely(f2fs_cp_error(sbi)))
-				return -EIO;
+				break;
 		}
 	}
 

@@ -1350,9 +1354,6 @@ static int do_checkpoint(struct f2fs_sb_info *sbi, struct cp_control *cpc)
 	/* wait for previous submitted meta pages writeback */
 	wait_on_all_pages_writeback(sbi);
 
-	if (unlikely(f2fs_cp_error(sbi)))
-		return -EIO;
-
 	/* flush all device cache */
 	err = f2fs_flush_device_cache(sbi);
 	if (err)

@@ -1364,9 +1365,6 @@ static int do_checkpoint(struct f2fs_sb_info *sbi, struct cp_control *cpc)
 
 	f2fs_release_ino_entry(sbi, false);
 
-	if (unlikely(f2fs_cp_error(sbi)))
-		return -EIO;
-
 	clear_sbi_flag(sbi, SBI_IS_DIRTY);
 	clear_sbi_flag(sbi, SBI_NEED_CP);
 	__set_cp_next_pack(sbi);

fs/f2fs/checkpoint.c +0 -0

 }
 
 /*

Quoted 92 lines · Jaegeuk Kim

It turns out losing meta pages in shutdown period makes f2fs very unstable so that I could see many unexpected error conditions.

Let's keep meta pages for fault injection and sudden power-off tests.

Signed-off-by Jaegeuk Kim <[email protected]>

---
 fs/f2fs/checkpoint.c | 28 +++++++++++++---------------
 1 file changed, 13 insertions(+), 15 deletions(-)

fs/f2fs/checkpoint.c +12 -14

index 9f1c96caebda..37ec0f16448e 100644
@@ -242,11 +242,8 @@ static int __f2fs_write_meta_page(struct page *page,
 
 	trace_f2fs_writepage(page, META);
 
-	if (unlikely(f2fs_cp_error(sbi))) {
-		dec_page_count(sbi, F2FS_DIRTY_META);
-		unlock_page(page);
-		return 0;
-	}
+	if (unlikely(f2fs_cp_error(sbi)))
+		goto redirty_out;
 	if (unlikely(is_sbi_flag_set(sbi, SBI_POR_DOING)))
 		goto redirty_out;
 	if (wbc->for_reclaim && page->index < GET_SUM_BLOCK(sbi, 0))

@@ -1129,6 +1126,9 @@ static void wait_on_all_pages_writeback(struct f2fs_sb_info *sbi)
 		if (!get_pages(sbi, F2FS_WB_CP_DATA))
 			break;
 
+		if (unlikely(f2fs_cp_error(sbi)))
+			break;
+
 		io_schedule_timeout(5*HZ);
 	}
 	finish_wait(&sbi->cp_wait, &wait);

@@ -1202,8 +1202,12 @@ static void commit_checkpoint(struct f2fs_sb_info *sbi,
 
 	/* writeout cp pack 2 page */
 	err = __f2fs_write_meta_page(page, &wbc, FS_CP_META_IO);
-	f2fs_bug_on(sbi, err);
+	if (unlikely(err && f2fs_cp_error(sbi))) {
+		f2fs_put_page(page, 1);
+		return;
+	}
 
+	f2fs_bug_on(sbi, err);
 	f2fs_put_page(page, 0);
 
 	/* submit checkpoint (with barrier if NOBARRIER is not set) */

@@ -1229,7 +1233,7 @@ static int do_checkpoint(struct f2fs_sb_info *sbi, struct cp_control *cpc)
 	while (get_pages(sbi, F2FS_DIRTY_META)) {
 		f2fs_sync_meta_pages(sbi, META, LONG_MAX, FS_CP_META_IO);
 		if (unlikely(f2fs_cp_error(sbi)))
-			return -EIO;
+			break;
 	}
 
 	/*

@@ -1309,7 +1313,7 @@ static int do_checkpoint(struct f2fs_sb_info *sbi, struct cp_control *cpc)
 			f2fs_sync_meta_pages(sbi, META, LONG_MAX,
 							FS_CP_META_IO);
 			if (unlikely(f2fs_cp_error(sbi)))
-				return -EIO;
+				break;
 		}
 	}
 

@@ -1350,9 +1354,6 @@ static int do_checkpoint(struct f2fs_sb_info *sbi, struct cp_control *cpc)
 	/* wait for previous submitted meta pages writeback */
 	wait_on_all_pages_writeback(sbi);
 
-	if (unlikely(f2fs_cp_error(sbi)))
-		return -EIO;
-
 	/* flush all device cache */
 	err = f2fs_flush_device_cache(sbi);
 	if (err)

@@ -1364,9 +1365,6 @@ static int do_checkpoint(struct f2fs_sb_info *sbi, struct cp_control *cpc)
 
 	f2fs_release_ino_entry(sbi, false);
 
-	if (unlikely(f2fs_cp_error(sbi)))
-		return -EIO;
-
 	clear_sbi_flag(sbi, SBI_IS_DIRTY);
 	clear_sbi_flag(sbi, SBI_NEED_CP);
 	__set_cp_next_pack(sbi);

@@ -1381,7 +1379,7 @@ static int do_checkpoint(struct f2fs_sb_info *sbi, struct cp_control *cpc)
 

fs/f2fs/checkpoint.c static int do_checkpoint(struct f2fs_sb_info *sbi, struct cp_control *cpc) · line 1382

Quoted 4 lines

fs/f2fs/checkpoint.c +1 -1

 	f2fs_bug_on(sbi, get_pages(sbi, F2FS_DIRTY_DENTS));
 
-	return 0;
+	return unlikely(f2fs_cp_error(sbi));

return unlikely(f2fs_cp_error(sbi)) ? -EIO : 0;

Quoted 4 lines

fs/f2fs/checkpoint.c +0 -0

 }
 
 /*

In order to prevent abusing atomic writes by abnormal users, we've added a threshold, 20% over memory footprint, which disallows further atomic writes. Previously, however, SQLite doesn't know the files became normal, so that it could write stale data and commit on revoked normal database file.

Once f2fs detects such the abnormal behavior, this patch simply disables all the atomic operations such as:
- write_begin() gives EINVAL to avoid stale data writes, and SQLite will call F2FS_IOC_ABORT_VOLATILE_WRITE to notify aborting the transaction,
- F2FS_IOC_START_ATOMIC_WRITE gives EINVAL for SQLite to fall back normal journal_mode,
- F2FS_IOC_COMMIT_ATOMIC_WRITE gives EINVAL, if the file was revoked, so that Framework retries to submit the transaction given propagated SQLite error.

Note that, this patch also turns off atomic operations, if foreground GC tries to move atomic files too frequently.

Signed-off-by Jaegeuk Kim <[email protected]>

---
 fs/f2fs/data.c    |  7 ++++---
 fs/f2fs/f2fs.h    |  4 ++--
 fs/f2fs/file.c    |  6 +++++-
 fs/f2fs/gc.c      |  4 +---
 fs/f2fs/segment.c | 21 +++++++++++----------
 5 files changed, 23 insertions(+), 19 deletions(-)

fs/f2fs/data.c +4 -3

index 5e53d210e222..c9e75aa61c24 100644
@@ -2247,8 +2247,9 @@ static int f2fs_write_begin(struct file *file, struct address_space *mapping,
 	trace_f2fs_write_begin(inode, pos, len, flags);
 
 	if (f2fs_is_atomic_file(inode) &&
-			!f2fs_available_free_memory(sbi, INMEM_PAGES)) {
-		err = -ENOMEM;
+			(is_sbi_flag_set(sbi, SBI_DISABLE_ATOMIC_WRITE) ||
+			!f2fs_available_free_memory(sbi, INMEM_PAGES))) {
+		err = -EINVAL;
 		drop_atomic = true;
 		goto fail;
 	}

@@ -2331,7 +2332,7 @@ static int f2fs_write_begin(struct file *file, struct address_space *mapping,
 	f2fs_put_page(page, 1);
 	f2fs_write_failed(mapping, pos + len);
 	if (drop_atomic)
-		f2fs_drop_inmem_pages_all(sbi, false);
+		f2fs_disable_atomic_write(sbi);
 	return err;
 }
 

fs/f2fs/f2fs.h +2 -2

index 4d8b1de83143..1d5c8d543eda 100644
@@ -621,7 +621,6 @@ enum {
 
 enum {
 	GC_FAILURE_PIN,
-	GC_FAILURE_ATOMIC,
 	MAX_GC_FAILURE
 };
 

@@ -1066,6 +1065,7 @@ enum {
 	SBI_POR_DOING,				/* recovery is doing or not */
 	SBI_NEED_SB_WRITE,			/* need to recover superblock */
 	SBI_NEED_CP,				/* need to checkpoint */
+	SBI_DISABLE_ATOMIC_WRITE,		/* turn off atomic write */
 };
 
 enum {

@@ -2833,7 +2833,7 @@ void f2fs_destroy_node_manager_caches(void);
  */
 bool f2fs_need_SSR(struct f2fs_sb_info *sbi);
 void f2fs_register_inmem_page(struct inode *inode, struct page *page);
-void f2fs_drop_inmem_pages_all(struct f2fs_sb_info *sbi, bool gc_failure);
+void f2fs_disable_atomic_write(struct f2fs_sb_info *sbi);
 void f2fs_drop_inmem_pages(struct inode *inode);
 void f2fs_drop_inmem_page(struct inode *inode, struct page *page);
 int f2fs_commit_inmem_pages(struct inode *inode);

fs/f2fs/file.c +5 -1

index 6880c6f78d58..b029a4ed3bb0 100644
@@ -1682,6 +1682,9 @@ static int f2fs_ioc_start_atomic_write(struct file *filp)
 	if (!S_ISREG(inode->i_mode))
 		return -EINVAL;
 
+	if (is_sbi_flag_set(F2FS_I_SB(inode), SBI_DISABLE_ATOMIC_WRITE))
+		return -EINVAL;
+
 	ret = mnt_want_write_file(filp);
 	if (ret)
 		return ret;

@@ -1750,7 +1753,6 @@ static int f2fs_ioc_commit_atomic_write(struct file *filp)
 		ret = f2fs_do_sync_file(filp, 0, LLONG_MAX, 0, true);
 		if (!ret) {
 			clear_inode_flag(inode, FI_ATOMIC_FILE);
-			F2FS_I(inode)->i_gc_failures[GC_FAILURE_ATOMIC] = 0;
 			stat_dec_atomic_write(inode);
 		}
 	} else {

@@ -1853,6 +1855,8 @@ static int f2fs_ioc_abort_volatile_write(struct file *filp)
 		ret = f2fs_do_sync_file(filp, 0, LLONG_MAX, 0, true);
 	}
 
+	clear_inode_flag(inode, FI_ATOMIC_REVOKE_REQUEST);
+
 	inode_unlock(inode);
 
 	mnt_drop_write_file(filp);

fs/f2fs/gc.c +1 -3

index 9093be6e7a7d..6d762f3cdfc7 100644
@@ -629,7 +629,6 @@ static void move_data_block(struct inode *inode, block_t bidx,
 		goto out;
 
 	if (f2fs_is_atomic_file(inode)) {
-		F2FS_I(inode)->i_gc_failures[GC_FAILURE_ATOMIC]++;
 		F2FS_I_SB(inode)->skipped_atomic_files[gc_type]++;
 		goto out;
 	}

@@ -745,7 +744,6 @@ static void move_data_page(struct inode *inode, block_t bidx, int gc_type,
 		goto out;
 
 	if (f2fs_is_atomic_file(inode)) {
-		F2FS_I(inode)->i_gc_failures[GC_FAILURE_ATOMIC]++;
 		F2FS_I_SB(inode)->skipped_atomic_files[gc_type]++;
 		goto out;
 	}

@@ -1100,7 +1098,7 @@ int f2fs_gc(struct f2fs_sb_info *sbi, bool sync,
 		if (has_not_enough_free_secs(sbi, sec_freed, 0)) {
 			if (skipped_round > MAX_SKIP_ATOMIC_COUNT &&
 				skipped_round * 2 >= round)
-				f2fs_drop_inmem_pages_all(sbi, true);
+				f2fs_disable_atomic_write(sbi);
 			segno = NULL_SEGNO;
 			goto gc_more;
 		}

fs/f2fs/segment.c +11 -10

index 9efce174c51a..05877a2f1894 100644
@@ -274,11 +274,14 @@ static int __revoke_inmem_pages(struct inode *inode,
 	return err;
 }
 
-void f2fs_drop_inmem_pages_all(struct f2fs_sb_info *sbi, bool gc_failure)
+void f2fs_disable_atomic_write(struct f2fs_sb_info *sbi)
 {
 	struct list_head *head = &sbi->inode_list[ATOMIC_FILE];
 	struct inode *inode;
 	struct f2fs_inode_info *fi;
+
+	/* just turn it off */
+	set_sbi_flag(sbi, SBI_DISABLE_ATOMIC_WRITE);
 next:
 	spin_lock(&sbi->inode_lock[ATOMIC_FILE]);
 	if (list_empty(head)) {

@@ -290,17 +293,16 @@ void f2fs_drop_inmem_pages_all(struct f2fs_sb_info *sbi, bool gc_failure)
 	spin_unlock(&sbi->inode_lock[ATOMIC_FILE]);
 
 	if (inode) {
-		if (gc_failure) {
-			if (fi->i_gc_failures[GC_FAILURE_ATOMIC])
-				goto drop;
-			goto skip;
+		inode_lock(inode);
+		/* need to check whether it was already revoked */
+		if (f2fs_is_atomic_file(inode)) {
+			f2fs_drop_inmem_pages(inode);
+			set_inode_flag(inode, FI_ATOMIC_REVOKE_REQUEST);
 		}
-drop:
-		set_inode_flag(inode, FI_ATOMIC_REVOKE_REQUEST);
-		f2fs_drop_inmem_pages(inode);
+		inode_unlock(inode);
 		iput(inode);
 	}
-skip:
+
 	congestion_wait(BLK_RW_ASYNC, HZ/50);
 	cond_resched();
 	goto next;

@@ -320,7 +322,6 @@ void f2fs_drop_inmem_pages(struct inode *inode)
 	mutex_unlock(&fi->inmem_lock);
 
 	clear_inode_flag(inode, FI_ATOMIC_FILE);
-	fi->i_gc_failures[GC_FAILURE_ATOMIC] = 0;
 	stat_dec_atomic_write(inode);
 }
 

Quoted 14 lines · Jaegeuk Kim

In order to prevent abusing atomic writes by abnormal users, we've added a threshold, 20% over memory footprint, which disallows further atomic writes. Previously, however, SQLite doesn't know the files became normal, so that it could write stale data and commit on revoked normal database file.

Once f2fs detects such the abnormal behavior, this patch simply disables all the atomic operations such as:
- write_begin() gives EINVAL to avoid stale data writes, and SQLite will call F2FS_IOC_ABORT_VOLATILE_WRITE to notify aborting the transaction,
- F2FS_IOC_START_ATOMIC_WRITE gives EINVAL for SQLite to fall back normal journal_mode,
- F2FS_IOC_COMMIT_ATOMIC_WRITE gives EINVAL, if the file was revoked, so that Framework retries to submit the transaction given propagated SQLite error.

Quoted 2 lines

Note that, this patch also turns off atomic operations, if foreground GC tries to move atomic files too frequently.

Well, how about just keeping original implementation: shutdown atomic write for those files which are really affect fggc? Since intention of the original behavior is targeting to abnormal atomic write usage, e.g. open atomic_write file for very long time, then fggc will be blocked each time when moving its block. So shutdown it, fggc will recover.

Thanks,

Quoted 178 lines

Signed-off-by Jaegeuk Kim <[email protected]>

---
 fs/f2fs/data.c    |  7 ++++---
 fs/f2fs/f2fs.h    |  4 ++--
 fs/f2fs/file.c    |  6 +++++-
 fs/f2fs/gc.c      |  4 +---
 fs/f2fs/segment.c | 21 +++++++++++----------
 5 files changed, 23 insertions(+), 19 deletions(-)

fs/f2fs/data.c +4 -3

index 5e53d210e222..c9e75aa61c24 100644
@@ -2247,8 +2247,9 @@ static int f2fs_write_begin(struct file *file, struct address_space *mapping,
 	trace_f2fs_write_begin(inode, pos, len, flags);
 
 	if (f2fs_is_atomic_file(inode) &&
-			!f2fs_available_free_memory(sbi, INMEM_PAGES)) {
-		err = -ENOMEM;
+			(is_sbi_flag_set(sbi, SBI_DISABLE_ATOMIC_WRITE) ||
+			!f2fs_available_free_memory(sbi, INMEM_PAGES))) {
+		err = -EINVAL;
 		drop_atomic = true;
 		goto fail;
 	}

@@ -2331,7 +2332,7 @@ static int f2fs_write_begin(struct file *file, struct address_space *mapping,
 	f2fs_put_page(page, 1);
 	f2fs_write_failed(mapping, pos + len);
 	if (drop_atomic)
-		f2fs_drop_inmem_pages_all(sbi, false);
+		f2fs_disable_atomic_write(sbi);
 	return err;
 }
 

fs/f2fs/f2fs.h +2 -2

index 4d8b1de83143..1d5c8d543eda 100644
@@ -621,7 +621,6 @@ enum {
 
 enum {
 	GC_FAILURE_PIN,
-	GC_FAILURE_ATOMIC,
 	MAX_GC_FAILURE
 };
 

@@ -1066,6 +1065,7 @@ enum {
 	SBI_POR_DOING,				/* recovery is doing or not */
 	SBI_NEED_SB_WRITE,			/* need to recover superblock */
 	SBI_NEED_CP,				/* need to checkpoint */
+	SBI_DISABLE_ATOMIC_WRITE,		/* turn off atomic write */
 };
 
 enum {

@@ -2833,7 +2833,7 @@ void f2fs_destroy_node_manager_caches(void);
  */
 bool f2fs_need_SSR(struct f2fs_sb_info *sbi);
 void f2fs_register_inmem_page(struct inode *inode, struct page *page);
-void f2fs_drop_inmem_pages_all(struct f2fs_sb_info *sbi, bool gc_failure);
+void f2fs_disable_atomic_write(struct f2fs_sb_info *sbi);
 void f2fs_drop_inmem_pages(struct inode *inode);
 void f2fs_drop_inmem_page(struct inode *inode, struct page *page);
 int f2fs_commit_inmem_pages(struct inode *inode);

fs/f2fs/file.c +5 -1

index 6880c6f78d58..b029a4ed3bb0 100644
@@ -1682,6 +1682,9 @@ static int f2fs_ioc_start_atomic_write(struct file *filp)
 	if (!S_ISREG(inode->i_mode))
 		return -EINVAL;
 
+	if (is_sbi_flag_set(F2FS_I_SB(inode), SBI_DISABLE_ATOMIC_WRITE))
+		return -EINVAL;
+
 	ret = mnt_want_write_file(filp);
 	if (ret)
 		return ret;

@@ -1750,7 +1753,6 @@ static int f2fs_ioc_commit_atomic_write(struct file *filp)
 		ret = f2fs_do_sync_file(filp, 0, LLONG_MAX, 0, true);
 		if (!ret) {
 			clear_inode_flag(inode, FI_ATOMIC_FILE);
-			F2FS_I(inode)->i_gc_failures[GC_FAILURE_ATOMIC] = 0;
 			stat_dec_atomic_write(inode);
 		}
 	} else {

@@ -1853,6 +1855,8 @@ static int f2fs_ioc_abort_volatile_write(struct file *filp)
 		ret = f2fs_do_sync_file(filp, 0, LLONG_MAX, 0, true);
 	}
 
+	clear_inode_flag(inode, FI_ATOMIC_REVOKE_REQUEST);
+
 	inode_unlock(inode);
 
 	mnt_drop_write_file(filp);

fs/f2fs/gc.c +1 -3

index 9093be6e7a7d..6d762f3cdfc7 100644
@@ -629,7 +629,6 @@ static void move_data_block(struct inode *inode, block_t bidx,
 		goto out;
 
 	if (f2fs_is_atomic_file(inode)) {
-		F2FS_I(inode)->i_gc_failures[GC_FAILURE_ATOMIC]++;
 		F2FS_I_SB(inode)->skipped_atomic_files[gc_type]++;
 		goto out;
 	}

@@ -745,7 +744,6 @@ static void move_data_page(struct inode *inode, block_t bidx, int gc_type,
 		goto out;
 
 	if (f2fs_is_atomic_file(inode)) {
-		F2FS_I(inode)->i_gc_failures[GC_FAILURE_ATOMIC]++;
 		F2FS_I_SB(inode)->skipped_atomic_files[gc_type]++;
 		goto out;
 	}

@@ -1100,7 +1098,7 @@ int f2fs_gc(struct f2fs_sb_info *sbi, bool sync,
 		if (has_not_enough_free_secs(sbi, sec_freed, 0)) {
 			if (skipped_round > MAX_SKIP_ATOMIC_COUNT &&
 				skipped_round * 2 >= round)
-				f2fs_drop_inmem_pages_all(sbi, true);
+				f2fs_disable_atomic_write(sbi);
 			segno = NULL_SEGNO;
 			goto gc_more;
 		}

fs/f2fs/segment.c +11 -10

index 9efce174c51a..05877a2f1894 100644
@@ -274,11 +274,14 @@ static int __revoke_inmem_pages(struct inode *inode,
 	return err;
 }
 
-void f2fs_drop_inmem_pages_all(struct f2fs_sb_info *sbi, bool gc_failure)
+void f2fs_disable_atomic_write(struct f2fs_sb_info *sbi)
 {
 	struct list_head *head = &sbi->inode_list[ATOMIC_FILE];
 	struct inode *inode;
 	struct f2fs_inode_info *fi;
+
+	/* just turn it off */
+	set_sbi_flag(sbi, SBI_DISABLE_ATOMIC_WRITE);
 next:
 	spin_lock(&sbi->inode_lock[ATOMIC_FILE]);
 	if (list_empty(head)) {

@@ -290,17 +293,16 @@ void f2fs_drop_inmem_pages_all(struct f2fs_sb_info *sbi, bool gc_failure)
 	spin_unlock(&sbi->inode_lock[ATOMIC_FILE]);
 
 	if (inode) {
-		if (gc_failure) {
-			if (fi->i_gc_failures[GC_FAILURE_ATOMIC])
-				goto drop;
-			goto skip;
+		inode_lock(inode);
+		/* need to check whether it was already revoked */
+		if (f2fs_is_atomic_file(inode)) {
+			f2fs_drop_inmem_pages(inode);
+			set_inode_flag(inode, FI_ATOMIC_REVOKE_REQUEST);
 		}
-drop:
-		set_inode_flag(inode, FI_ATOMIC_REVOKE_REQUEST);
-		f2fs_drop_inmem_pages(inode);
+		inode_unlock(inode);
 		iput(inode);
 	}
-skip:
+
 	congestion_wait(BLK_RW_ASYNC, HZ/50);
 	cond_resched();
 	goto next;

@@ -320,7 +322,6 @@ void f2fs_drop_inmem_pages(struct inode *inode)
 	mutex_unlock(&fi->inmem_lock);
 
 	clear_inode_flag(inode, FI_ATOMIC_FILE);
-	fi->i_gc_failures[GC_FAILURE_ATOMIC] = 0;
 	stat_dec_atomic_write(inode);
 }
 

Quoted 18 lines · Chao Yu

Quoted 16 lines · Jaegeuk Kim

In order to prevent abusing atomic writes by abnormal users, we've added a threshold, 20% over memory footprint, which disallows further atomic writes. Previously, however, SQLite doesn't know the files became normal, so that it could write stale data and commit on revoked normal database file.

Once f2fs detects such the abnormal behavior, this patch simply disables all the atomic operations such as:
- write_begin() gives EINVAL to avoid stale data writes, and SQLite will call F2FS_IOC_ABORT_VOLATILE_WRITE to notify aborting the transaction,
- F2FS_IOC_START_ATOMIC_WRITE gives EINVAL for SQLite to fall back normal journal_mode,
- F2FS_IOC_COMMIT_ATOMIC_WRITE gives EINVAL, if the file was revoked, so that Framework retries to submit the transaction given propagated SQLite error.

Note that, this patch also turns off atomic operations, if foreground GC tries to move atomic files too frequently.

Quoted 5 lines

Well, how about just keeping original implementation: shutdown atomic write for those files which are really affect fggc? Since intention of the original behavior is targeting to abnormal atomic write usage, e.g. open atomic_write file for very long time, then fggc will be blocked each time when moving its block. So shutdown it, fggc will recover.

The point here is stopping sqlite to keep going wrong data writes even after we already revoked blocks.

Thanks,

Quoted 175 lines

Thanks,

Quoted 172 lines

Signed-off-by Jaegeuk Kim <[email protected]>

---
 fs/f2fs/data.c    |  7 ++++---
 fs/f2fs/f2fs.h    |  4 ++--
 fs/f2fs/file.c    |  6 +++++-
 fs/f2fs/gc.c      |  4 +---
 fs/f2fs/segment.c | 21 +++++++++++----------
 5 files changed, 23 insertions(+), 19 deletions(-)

fs/f2fs/data.c +4 -3

index 5e53d210e222..c9e75aa61c24 100644
@@ -2247,8 +2247,9 @@ static int f2fs_write_begin(struct file *file, struct address_space *mapping,
 	trace_f2fs_write_begin(inode, pos, len, flags);
 
 	if (f2fs_is_atomic_file(inode) &&
-			!f2fs_available_free_memory(sbi, INMEM_PAGES)) {
-		err = -ENOMEM;
+			(is_sbi_flag_set(sbi, SBI_DISABLE_ATOMIC_WRITE) ||
+			!f2fs_available_free_memory(sbi, INMEM_PAGES))) {
+		err = -EINVAL;
 		drop_atomic = true;
 		goto fail;
 	}

@@ -2331,7 +2332,7 @@ static int f2fs_write_begin(struct file *file, struct address_space *mapping,
 	f2fs_put_page(page, 1);
 	f2fs_write_failed(mapping, pos + len);
 	if (drop_atomic)
-		f2fs_drop_inmem_pages_all(sbi, false);
+		f2fs_disable_atomic_write(sbi);
 	return err;
 }
 

fs/f2fs/f2fs.h +2 -2

index 4d8b1de83143..1d5c8d543eda 100644
@@ -621,7 +621,6 @@ enum {
 
 enum {
 	GC_FAILURE_PIN,
-	GC_FAILURE_ATOMIC,
 	MAX_GC_FAILURE
 };
 

@@ -1066,6 +1065,7 @@ enum {
 	SBI_POR_DOING,				/* recovery is doing or not */
 	SBI_NEED_SB_WRITE,			/* need to recover superblock */
 	SBI_NEED_CP,				/* need to checkpoint */
+	SBI_DISABLE_ATOMIC_WRITE,		/* turn off atomic write */
 };
 
 enum {

@@ -2833,7 +2833,7 @@ void f2fs_destroy_node_manager_caches(void);
  */
 bool f2fs_need_SSR(struct f2fs_sb_info *sbi);
 void f2fs_register_inmem_page(struct inode *inode, struct page *page);
-void f2fs_drop_inmem_pages_all(struct f2fs_sb_info *sbi, bool gc_failure);
+void f2fs_disable_atomic_write(struct f2fs_sb_info *sbi);
 void f2fs_drop_inmem_pages(struct inode *inode);
 void f2fs_drop_inmem_page(struct inode *inode, struct page *page);
 int f2fs_commit_inmem_pages(struct inode *inode);

fs/f2fs/file.c +5 -1

index 6880c6f78d58..b029a4ed3bb0 100644
@@ -1682,6 +1682,9 @@ static int f2fs_ioc_start_atomic_write(struct file *filp)
 	if (!S_ISREG(inode->i_mode))
 		return -EINVAL;
 
+	if (is_sbi_flag_set(F2FS_I_SB(inode), SBI_DISABLE_ATOMIC_WRITE))
+		return -EINVAL;
+
 	ret = mnt_want_write_file(filp);
 	if (ret)
 		return ret;

@@ -1750,7 +1753,6 @@ static int f2fs_ioc_commit_atomic_write(struct file *filp)
 		ret = f2fs_do_sync_file(filp, 0, LLONG_MAX, 0, true);
 		if (!ret) {
 			clear_inode_flag(inode, FI_ATOMIC_FILE);
-			F2FS_I(inode)->i_gc_failures[GC_FAILURE_ATOMIC] = 0;
 			stat_dec_atomic_write(inode);
 		}
 	} else {

@@ -1853,6 +1855,8 @@ static int f2fs_ioc_abort_volatile_write(struct file *filp)
 		ret = f2fs_do_sync_file(filp, 0, LLONG_MAX, 0, true);
 	}
 
+	clear_inode_flag(inode, FI_ATOMIC_REVOKE_REQUEST);
+
 	inode_unlock(inode);
 
 	mnt_drop_write_file(filp);

fs/f2fs/gc.c +1 -3

index 9093be6e7a7d..6d762f3cdfc7 100644
@@ -629,7 +629,6 @@ static void move_data_block(struct inode *inode, block_t bidx,
 		goto out;
 
 	if (f2fs_is_atomic_file(inode)) {
-		F2FS_I(inode)->i_gc_failures[GC_FAILURE_ATOMIC]++;
 		F2FS_I_SB(inode)->skipped_atomic_files[gc_type]++;
 		goto out;
 	}

@@ -745,7 +744,6 @@ static void move_data_page(struct inode *inode, block_t bidx, int gc_type,
 		goto out;
 
 	if (f2fs_is_atomic_file(inode)) {
-		F2FS_I(inode)->i_gc_failures[GC_FAILURE_ATOMIC]++;
 		F2FS_I_SB(inode)->skipped_atomic_files[gc_type]++;
 		goto out;
 	}

@@ -1100,7 +1098,7 @@ int f2fs_gc(struct f2fs_sb_info *sbi, bool sync,
 		if (has_not_enough_free_secs(sbi, sec_freed, 0)) {
 			if (skipped_round > MAX_SKIP_ATOMIC_COUNT &&
 				skipped_round * 2 >= round)
-				f2fs_drop_inmem_pages_all(sbi, true);
+				f2fs_disable_atomic_write(sbi);
 			segno = NULL_SEGNO;
 			goto gc_more;
 		}

fs/f2fs/segment.c +11 -9

index 9efce174c51a..05877a2f1894 100644
@@ -274,11 +274,14 @@ static int __revoke_inmem_pages(struct inode *inode,
 	return err;
 }
 
-void f2fs_drop_inmem_pages_all(struct f2fs_sb_info *sbi, bool gc_failure)
+void f2fs_disable_atomic_write(struct f2fs_sb_info *sbi)
 {
 	struct list_head *head = &sbi->inode_list[ATOMIC_FILE];
 	struct inode *inode;
 	struct f2fs_inode_info *fi;
+
+	/* just turn it off */
+	set_sbi_flag(sbi, SBI_DISABLE_ATOMIC_WRITE);
 next:
 	spin_lock(&sbi->inode_lock[ATOMIC_FILE]);
 	if (list_empty(head)) {

@@ -290,17 +293,16 @@ void f2fs_drop_inmem_pages_all(struct f2fs_sb_info *sbi, bool gc_failure)
 	spin_unlock(&sbi->inode_lock[ATOMIC_FILE]);
 
 	if (inode) {
-		if (gc_failure) {
-			if (fi->i_gc_failures[GC_FAILURE_ATOMIC])
-				goto drop;
-			goto skip;
+		inode_lock(inode);
+		/* need to check whether it was already revoked */
+		if (f2fs_is_atomic_file(inode)) {
+			f2fs_drop_inmem_pages(inode);
+			set_inode_flag(inode, FI_ATOMIC_REVOKE_REQUEST);
 		}
-drop:
-		set_inode_flag(inode, FI_ATOMIC_REVOKE_REQUEST);
-		f2fs_drop_inmem_pages(inode);
+		inode_unlock(inode);
 		iput(inode);
 	}
-skip:
+
 	congestion_wait(BLK_RW_ASYNC, HZ/50);
 	cond_resched();
 	goto next;

@@ -320,7 +322,6 @@ void f2fs_drop_inmem_pages(struct inode *inode)
 	mutex_unlock(&fi->inmem_lock);
 

Quoted 6 lines

fs/f2fs/segment.c void f2fs_drop_inmem_pages(struct inode *inode) · line 327

Quoted 6 lines

fs/f2fs/segment.c +0 -1

 	clear_inode_flag(inode, FI_ATOMIC_FILE);
-	fi->i_gc_failures[GC_FAILURE_ATOMIC] = 0;
 	stat_dec_atomic_write(inode);
 }
 

------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most engaging tech sites, Slashdot.org! http://sdm.link/slashdot
_______________________________________________
Linux-f2fs-devel mailing list [email protected]
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel

Quoted 25 lines · Jaegeuk Kim

Quoted 23 lines · Chao Yu

Quoted 14 lines · Jaegeuk Kim

In order to prevent abusing atomic writes by abnormal users, we've added a threshold, 20% over memory footprint, which disallows further atomic writes. Previously, however, SQLite doesn't know the files became normal, so that it could write stale data and commit on revoked normal database file.

Once f2fs detects such the abnormal behavior, this patch simply disables all the atomic operations such as:
- write_begin() gives EINVAL to avoid stale data writes, and SQLite will call F2FS_IOC_ABORT_VOLATILE_WRITE to notify aborting the transaction,
- F2FS_IOC_START_ATOMIC_WRITE gives EINVAL for SQLite to fall back normal journal_mode,
- F2FS_IOC_COMMIT_ATOMIC_WRITE gives EINVAL, if the file was revoked, so that Framework retries to submit the transaction given propagated SQLite error.

Quoted 2 lines

Note that, this patch also turns off atomic operations, if foreground GC tries to move atomic files too frequently.

Well, how about just keeping original implementation: shutdown atomic write for those files which are really affect fggc? Since intention of the original behavior is targeting to abnormal atomic write usage, e.g. open atomic_write file for very long time, then fggc will be blocked each time when moving its block. So shutdown it, fggc will recover.

Quoted 2 lines

The point here is stopping sqlite to keep going wrong data writes even after we already revoked blocks.

Yes, that's correct, what I mean is that if we can do that with smaller granularity like just revoke blocks for file which is blocking fggc, it will affect system/sqlite flow much less than forcing closing all atomic_write.

From 64d2becb82a496c2e2c04abeed42efa3b401ee20 Mon Sep 17 00:00:00 2001

Quoted 26 lines · Chao Yu

Quoted 25 lines · Jaegeuk Kim

Quoted 23 lines · Chao Yu

Quoted 14 lines · Jaegeuk Kim

In order to prevent abusing atomic writes by abnormal users, we've added a threshold, 20% over memory footprint, which disallows further atomic writes. Previously, however, SQLite doesn't know the files became normal, so that it could write stale data and commit on revoked normal database file.

Once f2fs detects such the abnormal behavior, this patch simply disables all the atomic operations such as:
- write_begin() gives EINVAL to avoid stale data writes, and SQLite will call F2FS_IOC_ABORT_VOLATILE_WRITE to notify aborting the transaction,
- F2FS_IOC_START_ATOMIC_WRITE gives EINVAL for SQLite to fall back normal journal_mode,
- F2FS_IOC_COMMIT_ATOMIC_WRITE gives EINVAL, if the file was revoked, so that Framework retries to submit the transaction given propagated SQLite error.

Quoted 2 lines

Note that, this patch also turns off atomic operations, if foreground GC tries to move atomic files too frequently.

Well, how about just keeping original implementation: shutdown atomic write for those files which are really affect fggc? Since intention of the original behavior is targeting to abnormal atomic write usage, e.g. open atomic_write file for very long time, then fggc will be blocked each time when moving its block. So shutdown it, fggc will recover.

Quoted 6 lines

Quoted 2 lines

The point here is stopping sqlite to keep going wrong data writes even after we already revoked blocks.

Yes, that's correct, what I mean is that if we can do that with smaller granularity like just revoke blocks for file which is blocking fggc, it will affect system/sqlite flow much less than forcing closing all atomic_write.

How about this?

From 64d2becb82a496c2e2c04abeed42efa3b401ee20 Mon Sep 17 00:00:00 2001
From: Jaegeuk Kim <[email protected]>
Date: Fri, 27 Jul 2018 18:15:11 +0900
Subject: [PATCH] f2fs: don't allow any writes on aborted atomic writes

In order to prevent abusing atomic writes by abnormal users, we've added a threshold, 20% over memory footprint, which disallows further atomic writes. Previously, however, SQLite doesn't know the files became normal, so that it could write stale data and commit on revoked normal database file.

Once f2fs detects such the abnormal behavior, this patch tries to avoid further writes in write_begin().

Signed-off-by Jaegeuk Kim <[email protected]>

---
 fs/f2fs/data.c | 3 ++-
 fs/f2fs/file.c | 7 ++++++-
 2 files changed, 8 insertions(+), 2 deletions(-)

fs/f2fs/data.c +2 -1

index 22dd00c6e241..02ec2603725f 100644
@@ -2295,7 +2295,8 @@ static int f2fs_write_begin(struct file *file, struct address_space *mapping,
 	trace_f2fs_write_begin(inode, pos, len, flags);
 
 	if (f2fs_is_atomic_file(inode) &&
-			!f2fs_available_free_memory(sbi, INMEM_PAGES)) {
+			(is_inode_flag_set(inode, FI_ATOMIC_REVOKE_REQUEST) ||
+			!f2fs_available_free_memory(sbi, INMEM_PAGES))) {
 		err = -ENOMEM;
 		drop_atomic = true;
 		goto fail;

fs/f2fs/file.c +6 -1

index ff2cb8fb6934..c2c47f3248c4 100644
@@ -1708,8 +1708,11 @@ static int f2fs_ioc_start_atomic_write(struct file *filp)
 
 	down_write(&F2FS_I(inode)->i_gc_rwsem[WRITE]);
 
-	if (f2fs_is_atomic_file(inode))
+	if (f2fs_is_atomic_file(inode)) {
+		if (is_inode_flag_set(inode, FI_ATOMIC_REVOKE_REQUEST))
+			ret = -EINVAL;
 		goto out;
+	}
 
 	ret = f2fs_convert_inline_inode(inode);
 	if (ret)

@@ -1871,6 +1874,8 @@ static int f2fs_ioc_abort_volatile_write(struct file *filp)
 		ret = f2fs_do_sync_file(filp, 0, LLONG_MAX, 0, true);
 	}
 
+	clear_inode_flag(inode, FI_ATOMIC_REVOKE_REQUEST);
+
 	inode_unlock(inode);
 
 	mnt_drop_write_file(filp);

Quoted 27 lines · Jaegeuk Kim

Quoted 26 lines · Chao Yu

Quoted 25 lines · Jaegeuk Kim

Quoted 23 lines · Chao Yu

Quoted 14 lines · Jaegeuk Kim

In order to prevent abusing atomic writes by abnormal users, we've added a threshold, 20% over memory footprint, which disallows further atomic writes. Previously, however, SQLite doesn't know the files became normal, so that it could write stale data and commit on revoked normal database file.

Once f2fs detects such the abnormal behavior, this patch simply disables all the atomic operations such as:
- write_begin() gives EINVAL to avoid stale data writes, and SQLite will call F2FS_IOC_ABORT_VOLATILE_WRITE to notify aborting the transaction,
- F2FS_IOC_START_ATOMIC_WRITE gives EINVAL for SQLite to fall back normal journal_mode,
- F2FS_IOC_COMMIT_ATOMIC_WRITE gives EINVAL, if the file was revoked, so that Framework retries to submit the transaction given propagated SQLite error.

Quoted 2 lines

Note that, this patch also turns off atomic operations, if foreground GC tries to move atomic files too frequently.

Well, how about just keeping original implementation: shutdown atomic write for those files which are really affect fggc? Since intention of the original behavior is targeting to abnormal atomic write usage, e.g. open atomic_write file for very long time, then fggc will be blocked each time when moving its block. So shutdown it, fggc will recover.

Quoted 8 lines

Quoted 6 lines

Quoted 2 lines

The point here is stopping sqlite to keep going wrong data writes even after we already revoked blocks.

Yes, that's correct, what I mean is that if we can do that with smaller granularity like just revoke blocks for file which is blocking fggc, it will affect system/sqlite flow much less than forcing closing all atomic_write.

How about this?

Yes, it looks good to me. :)

Quoted 11 lines

From 64d2becb82a496c2e2c04abeed42efa3b401ee20 Mon Sep 17 00:00:00 2001
From: Jaegeuk Kim <[email protected]>
Date: Fri, 27 Jul 2018 18:15:11 +0900
Subject: [PATCH] f2fs: don't allow any writes on aborted atomic writes

In order to prevent abusing atomic writes by abnormal users, we've added a threshold, 20% over memory footprint, which disallows further atomic writes. Previously, however, SQLite doesn't know the files became normal, so that it could write stale data and commit on revoked normal database file.

Quoted 4 lines

Once f2fs detects such the abnormal behavior, this patch tries to avoid further writes in write_begin().

Signed-off-by Jaegeuk Kim <[email protected]>

Reviewed-by Chao Yu <[email protected]>

Thanks,

Quoted 12 lines

---
 fs/f2fs/data.c | 3 ++-
 fs/f2fs/file.c | 7 ++++++-
 2 files changed, 8 insertions(+), 2 deletions(-)

fs/f2fs/data.c +0 -0

index 22dd00c6e241..02ec2603725f 100644
@@ -2295,7 +2295,8 @@ static int f2fs_write_begin(struct file *file, struct address_space *mapping,
 	trace_f2fs_write_begin(inode, pos, len, flags);
 

fs/f2fs/data.c static int f2fs_write_begin(struct file *file, struct address_space *mapping, · line 2299

Quoted 4 lines

fs/f2fs/data.c +2 -1

 	if (f2fs_is_atomic_file(inode) &&
-			!f2fs_available_free_memory(sbi, INMEM_PAGES)) {
+			(is_inode_flag_set(inode, FI_ATOMIC_REVOKE_REQUEST) ||
+			!f2fs_available_free_memory(sbi, INMEM_PAGES))) {

Quoted 30 lines

fs/f2fs/data.c +0 -0

 		err = -ENOMEM;
 		drop_atomic = true;
 		goto fail;

fs/f2fs/file.c +6 -1

index ff2cb8fb6934..c2c47f3248c4 100644
@@ -1708,8 +1708,11 @@ static int f2fs_ioc_start_atomic_write(struct file *filp)
 
 	down_write(&F2FS_I(inode)->i_gc_rwsem[WRITE]);
 
-	if (f2fs_is_atomic_file(inode))
+	if (f2fs_is_atomic_file(inode)) {
+		if (is_inode_flag_set(inode, FI_ATOMIC_REVOKE_REQUEST))
+			ret = -EINVAL;
 		goto out;
+	}
 
 	ret = f2fs_convert_inline_inode(inode);
 	if (ret)

@@ -1871,6 +1874,8 @@ static int f2fs_ioc_abort_volatile_write(struct file *filp)
 		ret = f2fs_do_sync_file(filp, 0, LLONG_MAX, 0, true);
 	}
 
+	clear_inode_flag(inode, FI_ATOMIC_REVOKE_REQUEST);
+
 	inode_unlock(inode);
 
 	mnt_drop_write_file(filp);