Messages in this thread | | | Subject | Re: [PATCH] nvmet: fix uninitialized work for zero kato | From | Sagi Grimberg <> | Date | Tue, 13 Oct 2020 15:18:51 -0700 |
| |
> Hit a warning: > WARNING: CPU: 1 PID: 241 at kernel/workqueue.c:1627 __queue_delayed_work+0x6d/0x90 > with trace: > mod_delayed_work_on+0x59/0x90 > nvmet_update_cc+0xee/0x100 [nvmet] > nvmet_execute_prop_set+0x72/0x80 [nvmet] > nvmet_tcp_try_recv_pdu+0x2f7/0x770 [nvmet_tcp] > nvmet_tcp_io_work+0x63f/0xb2d [nvmet_tcp] > ... > > This could be reproduced easily with a keep alive time 0: > nvme connect -t tcp -n NQN -a ADDR -s PORT --keep-alive-tmo=0 > > The reason is: > Starting an uninitialized work when initiator connects with zero > kato. Althrough keep-alive timer is disabled during allocating a ctrl > (fix in 0d3b6a8d213a), ka_work still has a chance to run > (called by nvmet_start_ctrl to detect dead host).
This should have a "Fixes:" tag.
> > Initilize ka_work during allocating ctrl, and set a reasonable kato > before scheduling ka_work. > > Signed-off-by: zhenwei pi <pizhenwei@bytedance.com> > --- > drivers/nvme/target/core.c | 16 +++++++++++----- > 1 file changed, 11 insertions(+), 5 deletions(-) > > diff --git a/drivers/nvme/target/core.c b/drivers/nvme/target/core.c > index b7b63330b5ef..3c5b2b065476 100644 > --- a/drivers/nvme/target/core.c > +++ b/drivers/nvme/target/core.c > @@ -19,6 +19,8 @@ struct workqueue_struct *buffered_io_wq; > static const struct nvmet_fabrics_ops *nvmet_transports[NVMF_TRTYPE_MAX]; > static DEFINE_IDA(cntlid_ida); > > +#define NVMET_DEFAULT_KATO 5 > + > /* > * This read/write semaphore is used to synchronize access to configuration > * information on a target system that will result in discovery log page > @@ -385,6 +387,11 @@ static void nvmet_keep_alive_timer(struct work_struct *work) > if (cmd_seen) { > pr_debug("ctrl %d reschedule traffic based keep-alive timer\n", > ctrl->cntlid); > + > + /* run once, trigger from nvmet_start_ctrl to detect dead link */ > + if (!ctrl->kato) > + return; > + > schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
It will be better to just schedule/mod the ka_work if kato != 0, other changes in the patch aren't needed IMO.
> return; > } > @@ -403,15 +410,11 @@ static void nvmet_start_keep_alive_timer(struct nvmet_ctrl *ctrl) > pr_debug("ctrl %d start keep-alive timer for %d secs\n", > ctrl->cntlid, ctrl->kato); > > - INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer); > schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ); > } > > static void nvmet_stop_keep_alive_timer(struct nvmet_ctrl *ctrl) > { > - if (unlikely(ctrl->kato == 0)) > - return; > - > pr_debug("ctrl %d stop keep-alive\n", ctrl->cntlid); > > cancel_delayed_work_sync(&ctrl->ka_work); > @@ -1107,6 +1110,8 @@ static inline u8 nvmet_cc_iocqes(u32 cc) > > static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl) > { > + u32 kato = ctrl->kato ? ctrl->kato : NVMET_DEFAULT_KATO; > +
The controller shouldn't have a default value, it should receive the desired value from the host.
> lockdep_assert_held(&ctrl->lock); > > if (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES || > @@ -1126,7 +1131,7 @@ static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl) > * in case a host died before it enabled the controller. Hence, simply > * reset the keep alive timer when the controller is enabled. > */ > - mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ); > + mod_delayed_work(system_wq, &ctrl->ka_work, kato * HZ); > } > > static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl) > @@ -1378,6 +1383,7 @@ u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn, > > /* keep-alive timeout in seconds */ > ctrl->kato = DIV_ROUND_UP(kato, 1000); > + INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer); > > ctrl->err_counter = 0; > spin_lock_init(&ctrl->error_lock); >
| |