The code in ocfs2_dio_end_io_write() estimates number of necessary transaction credits using ocfs2_calc_extend_credits(). This however does not take into account that the IO could be arbitrarily large and can contain arbitrary number of extents. Extent tree manipulations do often extend the current transaction but not in all of the cases. For example if we have only single block extents in the tree, ocfs2_mark_extent_written() will end up calling ocfs2_replace_extent_rec() all the time and we will never extend the current transaction and eventually exhaust all the transaction credits if the IO contains many single block extents. Once that happens a WARN_ON(jbd2_handle_buffer_credits(handle) <= 0) is triggered in jbd2_journal_dirty_metadata() and subsequently OCFS2 aborts in response to this error. This was actually triggered by one of our customers on a heavily fragmented OCFS2 filesystem.
To fix the issue make sure the transaction always has enough credits for one extent insert before each call of ocfs2_mark_extent_written().
CC: stable@vger.kernel.org Fixes: c15471f79506 ("ocfs2: fix sparse file & data ordering issue in direct io") Reviewed-by: Joseph Qi joseph.qi@linux.alibaba.com Reviewed-by: Heming Zhao heming.zhao@suse.com Signed-off-by: Jan Kara jack@suse.cz --- fs/ocfs2/aops.c | 5 +++++ fs/ocfs2/journal.c | 17 +++++++++++++++++ fs/ocfs2/journal.h | 2 ++ fs/ocfs2/ocfs2_trace.h | 2 ++ 4 files changed, 26 insertions(+)
Changes since v1: * Added Reviewed-by tags, Fixes tag, CC stable * Updated changelog to describe better the observed issues
diff --git a/fs/ocfs2/aops.c b/fs/ocfs2/aops.c index b82185075de7..469143110f31 100644 --- a/fs/ocfs2/aops.c +++ b/fs/ocfs2/aops.c @@ -2368,6 +2368,11 @@ static int ocfs2_dio_end_io_write(struct inode *inode, }
list_for_each_entry(ue, &dwc->dw_zero_list, ue_node) { + ret = ocfs2_assure_trans_credits(handle, credits); + if (ret < 0) { + mlog_errno(ret); + break; + } ret = ocfs2_mark_extent_written(inode, &et, handle, ue->ue_cpos, 1, ue->ue_phys, diff --git a/fs/ocfs2/journal.c b/fs/ocfs2/journal.c index 604fea3a26ff..4866d7d3ac57 100644 --- a/fs/ocfs2/journal.c +++ b/fs/ocfs2/journal.c @@ -445,6 +445,23 @@ int ocfs2_extend_trans(handle_t *handle, int nblocks) return status; }
+/* + * Make sure handle has at least 'nblocks' credits available. If it does not + * have that many credits available, we will try to extend the handle to have + * enough credits. If that fails, we will restart transaction to have enough + * credits. Similar notes regarding data consistency and locking implications + * as for ocfs2_extend_trans() apply here. + */ +int ocfs2_assure_trans_credits(handle_t *handle, int nblocks) +{ + int old_nblks = jbd2_handle_buffer_credits(handle); + + trace_ocfs2_assure_trans_credits(old_nblks); + if (old_nblks >= nblocks) + return 0; + return ocfs2_extend_trans(handle, nblocks - old_nblks); +} + /* * If we have fewer than thresh credits, extend by OCFS2_MAX_TRANS_DATA. * If that fails, restart the transaction & regain write access for the diff --git a/fs/ocfs2/journal.h b/fs/ocfs2/journal.h index 41c9fe7e62f9..e3c3a35dc5e0 100644 --- a/fs/ocfs2/journal.h +++ b/fs/ocfs2/journal.h @@ -243,6 +243,8 @@ handle_t *ocfs2_start_trans(struct ocfs2_super *osb, int ocfs2_commit_trans(struct ocfs2_super *osb, handle_t *handle); int ocfs2_extend_trans(handle_t *handle, int nblocks); +int ocfs2_assure_trans_credits(handle_t *handle, + int nblocks); int ocfs2_allocate_extend_trans(handle_t *handle, int thresh);
diff --git a/fs/ocfs2/ocfs2_trace.h b/fs/ocfs2/ocfs2_trace.h index 9898c11bdfa1..d194d202413d 100644 --- a/fs/ocfs2/ocfs2_trace.h +++ b/fs/ocfs2/ocfs2_trace.h @@ -2577,6 +2577,8 @@ DEFINE_OCFS2_ULL_UINT_EVENT(ocfs2_commit_cache_end);
DEFINE_OCFS2_INT_INT_EVENT(ocfs2_extend_trans);
+DEFINE_OCFS2_INT_EVENT(ocfs2_assure_trans_credits); + DEFINE_OCFS2_INT_EVENT(ocfs2_extend_trans_restart);
DEFINE_OCFS2_INT_INT_EVENT(ocfs2_allocate_extend_trans);