From ecf753b3701d1556cdb57bf8442caf05b1030be6 Mon Sep 17 00:00:00 2001 From: Chao Yu Date: Sat, 2 Feb 2019 17:33:01 +0800 Subject: f2fs: fix potential data inconsistence of checkpoint Previously, we changed lock from cp_rwsem to node_change, it solved the deadlock issue which was caused by below race condition: Thread A Thread B - f2fs_setattr - f2fs_lock_op -- read_lock - dquot_transfer - __dquot_transfer - dquot_acquire - commit_dqblk - f2fs_quota_write - f2fs_write_begin - f2fs_write_failed - write_checkpoint - block_operations - f2fs_lock_all -- write_lock - f2fs_truncate_blocks - f2fs_lock_op -- read_lock But it breaks the sematics of cp_rwsem, in other callers like: - f2fs_file_write_iter -> f2fs_write_begin -> f2fs_write_failed - f2fs_direct_IO -> f2fs_write_failed We allow to truncate dnode w/o cp_rwsem held, result in incorrect sit bitmap update, which can cause further data corruption. So this patch reverts previous fix implementation, and try to fix deadlock by skipping calling f2fs_truncate_blocks() in f2fs_write_failed() only for quota file, and keep the preallocated data/node in the tail of quota file, we can expecte that the preallocated space can be used to store quota info latter soon. Fixes: af033b2aa8a8 ("f2fs: guarantee journalled quota data by checkpoint") Signed-off-by: Gao Xiang Signed-off-by: Sheng Yong Signed-off-by: Chao Yu Signed-off-by: Jaegeuk Kim --- fs/f2fs/inline.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'fs/f2fs/inline.c') diff --git a/fs/f2fs/inline.c b/fs/f2fs/inline.c index 6b4aa5db6a9a..ec067935ca78 100644 --- a/fs/f2fs/inline.c +++ b/fs/f2fs/inline.c @@ -298,7 +298,7 @@ process_inline: clear_inode_flag(inode, FI_INLINE_DATA); f2fs_put_page(ipage, 1); } else if (ri && (ri->i_inline & F2FS_INLINE_DATA)) { - if (f2fs_truncate_blocks(inode, 0, false, false)) + if (f2fs_truncate_blocks(inode, 0, false)) return false; goto process_inline; } @@ -470,7 +470,7 @@ static int f2fs_add_inline_entries(struct inode *dir, void *inline_dentry) return 0; punch_dentry_pages: truncate_inode_pages(&dir->i_data, 0); - f2fs_truncate_blocks(dir, 0, false, false); + f2fs_truncate_blocks(dir, 0, false); f2fs_remove_dirty_inode(dir); return err; } -- cgit v1.2.3 From 557dadbd8114716f215d72d2d89060720e5ad45d Mon Sep 17 00:00:00 2001 From: Chao Yu Date: Tue, 12 Mar 2019 15:44:27 +0800 Subject: f2fs: fix to avoid deadlock in f2fs_read_inline_dir() As Jiqun Li reported in bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=202883 sometimes, dead lock when make system call SYS_getdents64 with fsync() is called by another process. monkey running on android9.0 1. task 9785 held sbi->cp_rwsem and waiting lock_page() 2. task 10349 held mm_sem and waiting sbi->cp_rwsem 3. task 9709 held lock_page() and waiting mm_sem so this is a dead lock scenario. task stack is show by crash tools as following crash_arm64> bt ffffffc03c354080 PID: 9785 TASK: ffffffc03c354080 CPU: 1 COMMAND: "RxIoScheduler-3" >> #7 [ffffffc01b50fac0] __lock_page at ffffff80081b11e8 crash-arm64> bt 10349 PID: 10349 TASK: ffffffc018b83080 CPU: 1 COMMAND: "BUGLY_ASYNC_UPL" >> #3 [ffffffc01f8cfa40] rwsem_down_read_failed at ffffff8008a93afc PC: 00000033 LR: 00000000 SP: 00000000 PSTATE: ffffffffffffffff crash-arm64> bt 9709 PID: 9709 TASK: ffffffc03e7f3080 CPU: 1 COMMAND: "IntentService[A" >> #3 [ffffffc001e67850] rwsem_down_read_failed at ffffff8008a93afc >> #8 [ffffffc001e67b80] el1_ia at ffffff8008084fc4 PC: ffffff8008274114 [compat_filldir64+120] LR: ffffff80083584d4 [f2fs_fill_dentries+448] SP: ffffffc001e67b80 PSTATE: 80400145 X29: ffffffc001e67b80 X28: 0000000000000000 X27: 000000000000001a X26: 00000000000093d7 X25: ffffffc070d52480 X24: 0000000000000008 X23: 0000000000000028 X22: 00000000d43dfd60 X21: ffffffc001e67e90 X20: 0000000000000011 X19: ffffff80093a4000 X18: 0000000000000000 X17: 0000000000000000 X16: 0000000000000000 X15: 0000000000000000 X14: ffffffffffffffff X13: 0000000000000008 X12: 0101010101010101 X11: 7f7f7f7f7f7f7f7f X10: 6a6a6a6a6a6a6a6a X9: 7f7f7f7f7f7f7f7f X8: 0000000080808000 X7: ffffff800827409c X6: 0000000080808000 X5: 0000000000000008 X4: 00000000000093d7 X3: 000000000000001a X2: 0000000000000011 X1: ffffffc070d52480 X0: 0000000000800238 >> #9 [ffffffc001e67be0] f2fs_fill_dentries at ffffff80083584d0 PC: 0000003c LR: 00000000 SP: 00000000 PSTATE: 000000d9 X12: f48a02ff X11: d4678960 X10: d43dfc00 X9: d4678ae4 X8: 00000058 X7: d4678994 X6: d43de800 X5: 000000d9 X4: d43dfc0c X3: d43dfc10 X2: d46799c8 X1: 00000000 X0: 00001068 Below potential deadlock will happen between three threads: Thread A Thread B Thread C - f2fs_do_sync_file - f2fs_write_checkpoint - down_write(&sbi->node_change) -- 1) - do_page_fault - down_write(&mm->mmap_sem) -- 2) - do_wp_page - f2fs_vm_page_mkwrite - getdents64 - f2fs_read_inline_dir - lock_page -- 3) - f2fs_sync_node_pages - lock_page -- 3) - __do_map_lock - down_read(&sbi->node_change) -- 1) - f2fs_fill_dentries - dir_emit - compat_filldir64 - do_page_fault - down_read(&mm->mmap_sem) -- 2) Since f2fs_readdir is protected by inode.i_rwsem, there should not be any updates in inode page, we're safe to lookup dents in inode page without its lock held, so taking off the lock to improve concurrency of readdir and avoid potential deadlock. Reported-by: Jiqun Li Signed-off-by: Chao Yu Signed-off-by: Jaegeuk Kim --- fs/f2fs/inline.c | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) (limited to 'fs/f2fs/inline.c') diff --git a/fs/f2fs/inline.c b/fs/f2fs/inline.c index ec067935ca78..8d3d0022cef8 100644 --- a/fs/f2fs/inline.c +++ b/fs/f2fs/inline.c @@ -659,6 +659,12 @@ int f2fs_read_inline_dir(struct file *file, struct dir_context *ctx, if (IS_ERR(ipage)) return PTR_ERR(ipage); + /* + * f2fs_readdir was protected by inode.i_rwsem, it is safe to access + * ipage without page's lock held. + */ + unlock_page(ipage); + inline_dentry = inline_data_addr(inode, ipage); make_dentry_ptr_inline(inode, &d, inline_dentry); @@ -667,7 +673,7 @@ int f2fs_read_inline_dir(struct file *file, struct dir_context *ctx, if (!err) ctx->pos = d.max; - f2fs_put_page(ipage, 1); + f2fs_put_page(ipage, 0); return err < 0 ? err : 0; } -- cgit v1.2.3