Changwei Ge
2017-Dec-28 06:38 UTC
[Ocfs2-devel] [PATCH v2] ocfs2: try to reuse extent block in dealloc without meta_alloc
Hi Alex, On 2017/12/28 14:22, alex chen wrote:> Hi Changwei, > >> @@ -1347,8 +1362,15 @@ static int ocfs2_shift_tree_depth(handle_t *handle, >> struct ocfs2_extent_list *root_el;>> @@ -1762,8 +1764,8 @@ int ocfs2_write_begin_nolock(struct address_space *mapping, >> */ >> ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), >> wc->w_di_bh); >> - ret = ocfs2_lock_allocators(inode, &et, >> - clusters_to_alloc, extents_to_split, >> + ret = ocfs2_lock_allocators(inode, &et, clusters_to_alloc, >> + 2*extents_to_split, >> &data_ac, &meta_ac); > We has doubled extent number in ocfs2_populate_write_desc(), so here we don't need to double again.I think your comment here makes sense. I will remove the multiplier 2 in my next version of this patch. Thanks, Changwei> > Thanks, > Alex>
John Lightsey
2017-Dec-29 21:15 UTC
[Ocfs2-devel] [PATCH v2] ocfs2: try to reuse extent block in dealloc without meta_alloc
On Thu, 2017-12-28 at 06:38 +0000, Changwei Ge wrote:> Hi Alex, > > On 2017/12/28 14:22, alex chen wrote: > > We has doubled extent number in ocfs2_populate_write_desc(), so > > here we don't need to double again. > > I think your comment here makes sense. > I will remove the multiplier 2 in my next version of this patch. >This may have made it function in testing where it should have failed. I backported this patch to Debian's 4.9.65 kernel today, removed the *2 multiplier, and hit a related assertion failure in this codepath: ocfs2_mark_extent_written ocfs2_change_extent_flag ocfs2_split_extent ocfs2_split_and_insert ocfs2_grow_tree ocfs2_add_branch ocfs2_create_new_meta_bhs ocfs2_claim_metadata The assertion failure in ocfs2_claim_metadata() was BUG_ON(ac->ac_bits_wanted < (ac->ac_bits_given + bits_wanted)); My guess is that meta_ac was allocated at the beginning of the write, then the tree was truncated during the write, so the actual allocation is requesting more than was originally reserved. It seems like the codepaths where meta_ac != NULL also need to know about the list of extents available in et_dealloc. The Perl code I used to hit this bug is attached. -------------- next part -------------- A non-text attachment was scrubbed... Name: directio.pl Type: application/x-perl Size: 2145 bytes Desc: not available Url : http://oss.oracle.com/pipermail/ocfs2-devel/attachments/20171229/396d8318/attachment.bin