From mboxrd@z Thu Jan 1 00:00:00 1970 From: "yukuai (C)" Subject: Re: [PATCH 4/4] block, bfq: update pos_root for idle bfq_queue in bfq_bfqq_move() Date: Wed, 22 Dec 2021 11:12:45 +0800 Message-ID: <6ca1e924-47fa-b94e-598c-69a9549eb68e@huawei.com> References: <20211221032135.878550-1-yukuai3@huawei.com> <20211221032135.878550-5-yukuai3@huawei.com> <20211221115001.GD24748@quack2.suse.cz> Mime-Version: 1.0 Content-Transfer-Encoding: quoted-printable Return-path: In-Reply-To: <20211221115001.GD24748-4I4JzKEfoa/jFM9bn6wA6Q@public.gmane.org> List-ID: Content-Type: text/plain; charset="iso-8859-1"; format="flowed" To: Jan Kara Cc: tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org, axboe-tSWWG44O7X1aa/9Udqfwiw@public.gmane.org, paolo.valente-QSEj5FYQhm4dnm+yROfE0A@public.gmane.org, fchecconi-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org, avanzini.arianna-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org, cgroups-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, linux-block-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, linux-kernel-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, yi.zhang-hv44wF8Li93QT0dZR+AlfA@public.gmane.org =D4=DA 2021/12/21 19:50, Jan Kara =D0=B4=B5=C0: > On Tue 21-12-21 11:21:35, Yu Kuai wrote: >> During code review, we found that if bfqq is not busy in >> bfq_bfqq_move(), bfq_pos_tree_add_move() won't be called for the bfqq, >> thus bfqq->pos_root still points to the old bfqg. However, the ref >> that bfqq hold for the old bfqg will be released, so it's possible >> that the old bfqg can be freed. This is problematic because the freed >> bfqg can still be accessed by bfqq->pos_root. >> >> Fix the problem by calling bfq_pos_tree_add_move() for idle bfqq >> as well. >> >> Fixes: e21b7a0b9887 ("block, bfq: add full hierarchical scheduling and c= groups support") >> Signed-off-by: Yu Kuai >=20 > I'm just wondering, how can it happen that !bfq_bfqq_busy() queue is in > pos_tree? Because bfq_remove_request() takes care to remove bfqq from the > pos_tree... Hi, It's right this is not a problem in common case. The problem seems to relate to queue merging and task migration. Because I once reporduced it with the same reporducer for the problem that offlined bfqg can be inserted into service tree. The uaf is exactly in bfq_remove_request->rb_rease(). However I didn't save the stack... I guess this is because bfq_del_bfqq_busy() is called from bfq_release_process_ref(), and queue merging prevert sunch bfqq to be freed, thus such bfqq is not in service tree, and it's pos_root can point to the old bfqg after bfq_bic_update_cgroup->bfq_bfqq_move. I haven't confirmed this, however, this patch itself only cleared bfqq->pos_root for idle bfqq, there should be no harm. Thanks, Kuai >=20 > Honza >=20 >> --- >> block/bfq-cgroup.c | 14 +++++++++----- >> 1 file changed, 9 insertions(+), 5 deletions(-) >> >> diff --git a/block/bfq-cgroup.c b/block/bfq-cgroup.c >> index 8e8cf6b3d946..822dd28ecf53 100644 >> --- a/block/bfq-cgroup.c >> +++ b/block/bfq-cgroup.c >> @@ -677,7 +677,6 @@ void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq= _queue *bfqq, >> bfq_deactivate_bfqq(bfqd, bfqq, false, false); >> else if (entity->on_st_or_in_serv) >> bfq_put_idle_entity(bfq_entity_service_tree(entity), entity); >> - bfqg_and_blkg_put(old_parent); >> =20 >> if (entity->parent && >> entity->parent->last_bfqq_created =3D=3D bfqq) >> @@ -690,11 +689,16 @@ void bfq_bfqq_move(struct bfq_data *bfqd, struct b= fq_queue *bfqq, >> /* pin down bfqg and its associated blkg */ >> bfqg_and_blkg_get(bfqg); >> =20 >> - if (bfq_bfqq_busy(bfqq)) { >> - if (unlikely(!bfqd->nonrot_with_queueing)) >> - bfq_pos_tree_add_move(bfqd, bfqq); >> + /* >> + * Don't leave the pos_root to old bfqg, since the ref to old bfqg will >> + * be released and the bfqg might be freed. >> + */ >> + if (unlikely(!bfqd->nonrot_with_queueing)) >> + bfq_pos_tree_add_move(bfqd, bfqq); >> + bfqg_and_blkg_put(old_parent); >> + >> + if (bfq_bfqq_busy(bfqq)) >> bfq_activate_bfqq(bfqd, bfqq); >> - } >> =20 >> if (!bfqd->in_service_queue && !bfqd->rq_in_driver) >> bfq_schedule_dispatch(bfqd); >> --=20 >> 2.31.1 >>