At Mon, 13 Jan 2014 17:09:45 +0800, Liu Yuan wrote: > > On Sun, Jan 12, 2014 at 08:12:02PM +0900, Hitoshi Mitake wrote: > > At Sun, 12 Jan 2014 18:39:29 +0800, > > Liu Yuan wrote: > > > > > > On Sun, Jan 12, 2014 at 12:47:14AM +0900, Hitoshi Mitake wrote: > > > > At Sat, 11 Jan 2014 23:06:22 +0800, > > > > Liu Yuan wrote: > > > > > > > > > > On Sun, Jan 12, 2014 at 12:02:39AM +0900, Hitoshi Mitake wrote: > > > > > > At Sat, 11 Jan 2014 22:58:24 +0800, > > > > > > Liu Yuan wrote: > > > > > > > > > > > > > > On Sat, Jan 11, 2014 at 11:44:29PM +0900, Hitoshi Mitake wrote: > > > > > > > > From: Hitoshi Mitake <mitake.hitoshi at lab.ntt.co.jp> > > > > > > > > > > > > > > > > Current sheepdog allows the condition of a number of zones < a number > > > > > > > > of copies. But "dog vdi check" cannot handle the case well because it > > > > > > > > tries to create copies of lost/corrupted objects based on > > > > > > > > inode->nr_copies. > > > > > > > > > > > > > > > > This patch lets "dog vdi check" create a minimum number of copies and > > > > > > > > zones. > > > > > > > > > > > > > > > > Reported-by: Marcin Mirosław <marcin at mejor.pl> > > > > > > > > Signed-off-by: Hitoshi Mitake <mitake.hitoshi at lab.ntt.co.jp> > > > > > > > > --- > > > > > > > > > > > > > > > > v2: renameing, zones_nr -> sd_zones_nr > > > > > > > > > > > > > > > > dog/vdi.c | 19 ++++++++++++------- > > > > > > > > 1 file changed, 12 insertions(+), 7 deletions(-) > > > > > > > > > > > > > > > > diff --git a/dog/vdi.c b/dog/vdi.c > > > > > > > > index 8fd4664..4847350 100644 > > > > > > > > --- a/dog/vdi.c > > > > > > > > +++ b/dog/vdi.c > > > > > > > > @@ -1800,11 +1800,11 @@ static void vdi_check_object_main(struct work *work) > > > > > > > > } > > > > > > > > > > > > > > > > static void queue_vdi_check_work(const struct sd_inode *inode, uint64_t oid, > > > > > > > > - uint64_t *done, struct work_queue *wq) > > > > > > > > + uint64_t *done, struct work_queue *wq, > > > > > > > > + int nr_copies) > > > > > > > > { > > > > > > > > struct vdi_check_info *info; > > > > > > > > const struct sd_vnode *tgt_vnodes[SD_MAX_COPIES]; > > > > > > > > - int nr_copies = inode->nr_copies; > > > > > > > > > > > > > > > > info = xzalloc(sizeof(*info) + sizeof(info->vcw[0]) * nr_copies); > > > > > > > > info->oid = oid; > > > > > > > > @@ -1830,6 +1830,7 @@ struct check_arg { > > > > > > > > const struct sd_inode *inode; > > > > > > > > uint64_t *done; > > > > > > > > struct work_queue *wq; > > > > > > > > + int nr_copies; > > > > > > > > }; > > > > > > > > > > > > > > > > static void check_cb(void *data, enum btree_node_type type, void *arg) > > > > > > > > @@ -1844,7 +1845,8 @@ static void check_cb(void *data, enum btree_node_type type, void *arg) > > > > > > > > oid = vid_to_data_oid(ext->vdi_id, ext->idx); > > > > > > > > *(carg->done) = (uint64_t)ext->idx * SD_DATA_OBJ_SIZE; > > > > > > > > vdi_show_progress(*(carg->done), carg->inode->vdi_size); > > > > > > > > - queue_vdi_check_work(carg->inode, oid, NULL, carg->wq); > > > > > > > > + queue_vdi_check_work(carg->inode, oid, NULL, carg->wq, > > > > > > > > + carg->nr_copies); > > > > > > > > } > > > > > > > > } > > > > > > > > } > > > > > > > > @@ -1855,8 +1857,9 @@ int do_vdi_check(const struct sd_inode *inode) > > > > > > > > uint64_t done = 0, oid; > > > > > > > > uint32_t vid; > > > > > > > > struct work_queue *wq; > > > > > > > > + int nr_copies = min((int)inode->nr_copies, sd_zones_nr); > > > > > > > > > > > > > > > > - if (sd_nodes_nr < inode->nr_copies) { > > > > > > > > + if (sd_nodes_nr < nr_copies) { > > > > > > > > > > > > > > I think we should check (sd_zones_nr < nr_copies) here > > > > > > > > > > > > The comparison is already done in the above min(). Why should we check > > > > > > it here? > > > > > > > > > > then we should remove this check since it will never happen or we can leave it a > > > > > safe guard here. Either way, since zone is the true entity, not 'node', so what > > > > > we should really abort here is the condition that zones_nr < copies, no? > > > > > > > > The condition sd_nodes_nr < nr_copies can be true when a number of > > > > living nodes is less than zones or copies. In such a case, we cannot > > > > recover objects and the panic problem rises. > > > > > > sd_nodes_nr is always >= sd_zones_nr, no? We use zone to indicate a sheep entity > > > for data distribution, not node(this is what your patch actually try to solve). > > > So we should panic out only when sd_zones_nr < nr_copies (for now this should > > > not happen, but we can have this for safe guard of code). > > > > sd_zones_nr < nr_copies can happen because the past "unsafe" mode is > > default now. And Marcin's situation was it. > > > > After a second thought, I think we don't need this patch and instead, we should > simply handle it as before: > > - abort when sd_zones_nr < nr_copies as we did it before. > - it is very difficult to handle 'vdi check' for erasure code when zones_nr < nr_copies I can add a conditional branch based on copy_policy for erasure coded vdis, but > - it is less meaningful to recovery/rebuild the nodes when zones_nr < nr_copies This is cleary *your own personal* opinion. Are you really thinking like this? I cannot accept it completely. It is clearly conflicting with current policy of object placement. Even if a number of zones is less than a number of copies, sheep should recover objects if it is required. Thanks, Hitoshi |