On Wed, 2018-04-11 at 13:02 -0700, tj@xxxxxxxxxx wrote: > On Wed, Apr 11, 2018 at 08:00:29PM +0000, Bart Van Assche wrote: > > On Wed, 2018-04-11 at 12:57 -0700, tj@xxxxxxxxxx wrote: > > > On Wed, Apr 11, 2018 at 01:55:25PM -0600, Bart Van Assche wrote: > > > > On 04/11/18 13:00, Alexandru Moise wrote: > > > > > But the root cause of it is in blkcg_init_queue() when blkg_create() returns > > > > > an ERR ptr, because it tries to insert into a populated index into blkcg->blkg_tree, > > > > > the entry that we fail to remove at __blk_release_queue(). > > > > > > > > Hello Alex, > > > > > > > > Had you considered something like the untested patch below? > > > > > > But queue init shouldn't fail here, right? > > > > Hello Tejun, > > > > Your question is not entirely clear to me. Are you referring to the atomic > > allocations in blkg_create() or are you perhaps referring to something else? > > Hmm.. maybe I'm confused but I thought that the fact that > blkcg_init_queue() fails itself is already a bug, which happens > because a previously destroyed queue left behind blkgs. Hello Tejun, I had missed the start of this thread so I was not aware of which problem Alex was trying to solve. In the description of v1 of this patch I read that Alex thinks that he ran into a scenario in which blk_queue_alloc_node() assigns a q->id that is still in use by another request queue? That's weird. The following code still occurs in __blk_release_queue(): ida_simple_remove(&blk_queue_ida, q->id); It's not clear to me how that remove call could happen *before* q->id is removed from the blkcg radix tree. Bart.
