| Message ID | 20141217133331.282ffc6f592fb329f3b7edad@linux-foundation.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Wed, Dec 17, 2014 at 01:33:31PM -0800, Andrew Morton wrote: > > So I now have a mess on my hands due to reordering > ocfs2-fix-journal-commit-deadlock.patch ahead of this patch. > > It concerns the label "out:". Should it be placed before or after the > call to ocfs2_unlock_pages()? > > My current copy of ocfs2_write_end_nolock() is below, followed by my > current version of > ocfs2-call-ocfs2_journal_access_di-before-ocfs2_journal_dirty-in-ocfs2_write_end_nolock.patch You want "out:" after ocfs2_unlock_pages() to give us a chance to free any locked pages on the write contesxt. Btw, I have the following notes for this patch: Putting the journal_access_di in ocfs2_write_end is the correct thing to do, thanks. I think we want to keep the journal_access_di in ocfs2_write_begin though as we may change the disk inode when marking unwritten extents (see the call to ocfs2_mark_extent_written()). So: - I would remove the comment above journal_access_di in write_begin but not the actual call as we may dirty the inode buffer later. - Move the call to journal_access_di to the top of ocfs2_write_end_nolock as I believe you might be missing some inode buffer updates there too. Thanks Andrew, --Mark -- Mark Fasheh
On 2014/12/18 7:00, Mark Fasheh wrote: > On Wed, Dec 17, 2014 at 01:33:31PM -0800, Andrew Morton wrote: >> >> So I now have a mess on my hands due to reordering >> ocfs2-fix-journal-commit-deadlock.patch ahead of this patch. >> >> It concerns the label "out:". Should it be placed before or after the >> call to ocfs2_unlock_pages()? >> >> My current copy of ocfs2_write_end_nolock() is below, followed by my >> current version of >> ocfs2-call-ocfs2_journal_access_di-before-ocfs2_journal_dirty-in-ocfs2_write_end_nolock.patch > > You want "out:" after ocfs2_unlock_pages() to give us a chance to free any > locked pages on the write contesxt. > > Btw, I have the following notes for this patch: > > > Putting the journal_access_di in ocfs2_write_end is the correct thing to do, > thanks. I think we want to keep the journal_access_di in ocfs2_write_begin > though as we may change the disk inode when marking unwritten extents > (see the call to ocfs2_mark_extent_written()). So: > > - I would remove the comment above journal_access_di in write_begin but not > the actual call as we may dirty the inode buffer later. > Hi, Mark, About this patch, do you mean that: keep the journal_access_di() in ocfs2_write_begin, and call journal_access_di() in the top ocfs2_write_end() again? But I don't think it as a good idea. In some scenario, jbd2_journal_restart() might be called after we call ocfs2_journal_access_di. and jbd2_journal_commit_transaction() will commit the transaction. So calling ocfs2_journal_access_di() in ocfs2_write_end() will lead to buffer_uptodate(bh) == 0, so BUG. Am I right? Thanks, yangwenfang > - Move the call to journal_access_di to the top of ocfs2_write_end_nolock as > I believe you might be missing some inode buffer updates there too. > > > Thanks Andrew, > --Mark > > -- > Mark Fasheh > > . >
diff -puN fs/ocfs2/aops.c~ocfs2-call-ocfs2_journal_access_di-before-ocfs2_journal_dirty-in-ocfs2_write_end_nolock fs/ocfs2/aops.c --- a/fs/ocfs2/aops.c~ocfs2-call-ocfs2_journal_access_di-before-ocfs2_journal_dirty-in-ocfs2_write_end_nolock +++ a/fs/ocfs2/aops.c @@ -1822,16 +1822,6 @@ try_again: if (ret) goto out_commit; } - /* - * We don't want this to fail in ocfs2_write_end(), so do it - * here. - */ - ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh, - OCFS2_JOURNAL_ACCESS_WRITE); - if (ret) { - mlog_errno(ret); - goto out_quota; - } /* * Fill our page array first. That way we've grabbed enough so @@ -1982,7 +1972,7 @@ int ocfs2_write_end_nolock(struct addres loff_t pos, unsigned len, unsigned copied, struct page *page, void *fsdata) { - int i; + int i, ret; unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1); struct inode *inode = mapping->host; struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); @@ -2032,6 +2022,14 @@ int ocfs2_write_end_nolock(struct addres } } + ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh, + OCFS2_JOURNAL_ACCESS_WRITE); + if (ret) { + copied = ret; + mlog_errno(ret); + goto out; + } + out_write_size: pos += copied; if (pos > i_size_read(inode)) { @@ -2053,6 +2051,7 @@ out_write_size: */ ocfs2_unlock_pages(wc); +out: ocfs2_commit_trans(osb, handle); ocfs2_run_deallocs(osb, &wc->w_dealloc);
So I now have a mess on my hands due to reordering ocfs2-fix-journal-commit-deadlock.patch ahead of this patch. It concerns the label "out:". Should it be placed before or after the call to ocfs2_unlock_pages()? My current copy of ocfs2_write_end_nolock() is below, followed by my current version of ocfs2-call-ocfs2_journal_access_di-before-ocfs2_journal_dirty-in-ocfs2_write_end_nolock.patch Thanks. int ocfs2_write_end_nolock(struct address_space *mapping, loff_t pos, unsigned len, unsigned copied, struct page *page, void *fsdata) { int i, ret; unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1); struct inode *inode = mapping->host; struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); struct ocfs2_write_ctxt *wc = fsdata; struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data; handle_t *handle = wc->w_handle; struct page *tmppage; if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) { ocfs2_write_end_inline(inode, pos, len, &copied, di, wc); goto out_write_size; } if (unlikely(copied < len)) { if (!PageUptodate(wc->w_target_page)) copied = 0; ocfs2_zero_new_buffers(wc->w_target_page, start+copied, start+len); } flush_dcache_page(wc->w_target_page); for(i = 0; i < wc->w_num_pages; i++) { tmppage = wc->w_pages[i]; if (tmppage == wc->w_target_page) { from = wc->w_target_from; to = wc->w_target_to; BUG_ON(from > PAGE_CACHE_SIZE || to > PAGE_CACHE_SIZE || to < from); } else { /* * Pages adjacent to the target (if any) imply * a hole-filling write in which case we want * to flush their entire range. */ from = 0; to = PAGE_CACHE_SIZE; } if (page_has_buffers(tmppage)) { if (ocfs2_should_order_data(inode)) ocfs2_jbd2_file_inode(wc->w_handle, inode); block_commit_write(tmppage, from, to); } } ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh, OCFS2_JOURNAL_ACCESS_WRITE); if (ret) { copied = ret; mlog_errno(ret); goto out; } out_write_size: pos += copied; if (pos > i_size_read(inode)) { i_size_write(inode, pos); mark_inode_dirty(inode); } inode->i_blocks = ocfs2_inode_sector_count(inode); di->i_size = cpu_to_le64((u64)i_size_read(inode)); inode->i_mtime = inode->i_ctime = CURRENT_TIME; di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec); di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec); ocfs2_update_inode_fsync_trans(handle, inode, 1); ocfs2_journal_dirty(handle, wc->w_di_bh); /* unlock pages before dealloc since it needs acquiring j_trans_barrier * lock, or it will cause a deadlock since journal commit threads holds * this lock and will ask for the page lock when flushing the data. * put it here to preserve the unlock order. */ ocfs2_unlock_pages(wc); out: ocfs2_commit_trans(osb, handle); ocfs2_run_deallocs(osb, &wc->w_dealloc); brelse(wc->w_di_bh); kfree(wc); return copied; } From: yangwenfang <vicky.yangwenfang@huawei.com> Subject: ocfs2: call ocfs2_journal_access_di() before ocfs2_journal_dirty() in ocfs2_write_end_nolock() After we call ocfs2_journal_access_di() in ocfs2_write_begin(), jbd2_journal_restart() may also be called, in this function transaction A's t_updates-- and obtains a new transaction B. If jbd2_journal_commit_transaction() is happened to commit transaction A, when t_updates==0, it will continue to complete commit and unfile buffer. So when jbd2_journal_dirty_metadata(), the handle is pointed a new transaction B, and the buffer head's journal head is already freed, jh->b_transaction == NULL, jh->b_next_transaction == NULL, it returns EINVAL, So it triggers the BUG_ON(status). thread 1: jbd2: ocfs2_write_begin jbd2_journal_commit_transaction ocfs2_write_begin_nolock ocfs2_start_trans jbd2__journal_start(t_updates+1, transaction A) ocfs2_journal_access_di ocfs2_write_cluster_by_desc ocfs2_mark_extent_written ocfs2_change_extent_flag ocfs2_split_extent ocfs2_extend_rotate_transaction jbd2_journal_restart (t_updates-1,transaction B) t_updates==0 __jbd2_journal_refile_buffer ocfs2_write_end ocfs2_write_end_nolock ocfs2_journal_dirty jbd2_journal_dirty_metadata(bug) ocfs2_commit_trans In ext4, I found that: jbd2_journal_get_write_access() called by ext4_write_end. ext4_write_begin ext4_journal_start __ext4_journal_start_sb ext4_journal_check_start jbd2__journal_start ext4_write_end ext4_mark_inode_dirty ext4_reserve_inode_write ext4_journal_get_write_access jbd2_journal_get_write_access ext4_mark_iloc_dirty ext4_do_update_inode ext4_handle_dirty_metadata jbd2_journal_dirty_metadata So I think we should put ocfs2_journal_access_di before ocfs2_journal_dirty in the ocfs2_write_end. and it works well after my modification. Signed-off-by: vicky <vicky.yangwenfang@huawei.com> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> --- fs/ocfs2/aops.c | 21 ++++++++++----------- 1 file changed, 10 insertions(+), 11 deletions(-)