RE: [PATCH 6/7] ipmi: Handle queued messages more certainly on panic

From: æåèå / KAWAIïHIDEHIRO
Date: Mon Aug 17 2015 - 22:00:01 EST


Hello Corey,

> -----Original Message-----
> From: Corey Minyard [mailto:tcminyard@xxxxxxxxx] On Behalf Of Corey Minyard
> Sent: Wednesday, August 12, 2015 1:13 PM
> To: æåèå / KAWAIïHIDEHIRO
> Cc: openipmi-developer@xxxxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx
> Subject: Re: [PATCH 6/7] ipmi: Handle queued messages more certainly on panic
>
> On 07/27/2015 12:55 AM, Hidehiro Kawai wrote:
> > panic_event() called as a panic notifier tries to flush queued
> > messages, but it can't handle them if the kernel panic happens
> > while processing a message. What happens depends on when the
> > kernel panics.
>
> Sorry this took so long, I've been traveling.

No problem.

> I have queued the patches before this one. They all look good and
> necessary.

Thank you for reviewing!

> I'm not so sure about this patch. It looks like the only thing that is
> a real issue is #2 below.
> It's not so important to avoid dropping messages.

Initially I thought dropping middle of queued messages breaks
some consistencies if a message depends on the preceding dropped
message. However, userland tools normally issue request messages
in sequential manner, so the above situation wouldn't happen.
Now, I think dropping a message is OK.

> Can this be simplified somehow to work around the issue at panic time if
> intf->curr_msg is set and smi_info->waiting_msg is not?

There are two cases where intf->curr_msg is set and
smi_info->waiting_msg is not; one is before (2) and the other
is after (3). If we decide to drop intf->curr_msg in both cases,
I can simplify this patch somewhat.

Regards,

Hidehiro Kawai
Hitachi, Ltd. Research & Development Group

>
> Thank you,
>
> -corey
>
> >
> > Here is the summary of message sending process.
> >
> > smi_send()
> > smi_add_send_msg()
> > (1) intf->curr_msg = msg
> > sender()
> > (2) smi_info->waiting_msg = msg
> >
> > <asynchronously called>
> > check_start_timer_thread()
> > start_next_msg()
> > smi_info->curr_msg = smi_info->waiting_msg
> > (3) smi_info->waiting_msg = NULL
> > (4) smi_info->handlers->start_transaction()
> >
> > <asynchronously called>
> > smi_event_handler()
> > (5) handle_transaction_done()
> > smi_info->curr_msg = NULL
> > deliver_recv_msg()
> > ipmi_smi_msg_received()
> > intf->curr_msg = NULL
> >
> > If the kernel panics before (1), the requested message will be
> > lost. But it can't be helped.
> >
> > If the kernel panics before (2), new message sent by
> > send_panic_events() is queued to intf->xmit_msgs because
> > intf->curr_msg is non-NULL. But the new message will be never
> > sent because no one sends intf->curr_msg. As the result, the
> > kernel hangs up.
> >
> > If the kernel panics before (3), intf->curr_msg will be sent by
> > set_run_to_completion(). It's no problem.
> >
> > If the kernel panics before (4), intf->curr_msg will be lost.
> > However, messages on intf->xmit_msgs will be handled.
> >
> > If the kernel panics before (5), we try to continue running the
> > state machine. It may successfully complete.
> >
> > If the kernel panics after (5), we will miss the response message
> > handling, but it's not much problem in the panic context.
> >
> > This patch tries to handle messages in intf->curr_msg and
> > intf->xmit_msgs only once without losing them. To achieve this,
> > this patch does that:
> > - if a message is in intf->curr_msg or intf->xmit_msgs and
> > start_transaction() for the message hasn't been done yet,
> > resend it
> > - if start_transaction() for a message has been called,
> > just continue to run the state machine
> > - if the transaction has been completed, do nothing
> >
> > >From the perspective of implementation, these are done by keeping
> > smi_info->waiting_msg until start_transaction() is completed and
> > by keeping new flag IPMI_MSG_RESEND_ON_PANIC just before starting
> > the state machine.
> >
> > Signed-off-by: Hidehiro Kawai <hidehiro.kawai.ez@xxxxxxxxxxx>
> > ---
> > drivers/char/ipmi/ipmi_msghandler.c | 36 +++++++++++++++++++++++++++++++++++
> > drivers/char/ipmi/ipmi_si_intf.c | 5 ++++-
> > include/linux/ipmi_smi.h | 5 +++++
> > 3 files changed, 45 insertions(+), 1 deletion(-)
> >
> > diff --git a/drivers/char/ipmi/ipmi_msghandler.c b/drivers/char/ipmi/ipmi_msghandler.c
> > index 5a2d9fe..3dcd814 100644
> > --- a/drivers/char/ipmi/ipmi_msghandler.c
> > +++ b/drivers/char/ipmi/ipmi_msghandler.c
> > @@ -1493,6 +1493,8 @@ static struct ipmi_smi_msg *smi_add_send_msg(ipmi_smi_t intf,
> > struct ipmi_smi_msg *smi_msg,
> > int priority)
> > {
> > + smi_msg->flags |= IPMI_MSG_RESEND_ON_PANIC;
> > +
> > if (intf->curr_msg) {
> > if (priority > 0)
> > list_add_tail(&smi_msg->link, &intf->hp_xmit_msgs);
> > @@ -4223,6 +4225,7 @@ struct ipmi_smi_msg *ipmi_alloc_smi_msg(void)
> > rv->done = free_smi_msg;
> > rv->user_data = NULL;
> > atomic_inc(&smi_msg_inuse_count);
> > + rv->flags = 0;
> > }
> > return rv;
> > }
> > @@ -4531,7 +4534,40 @@ static int panic_event(struct notifier_block *this,
> > spin_unlock(&intf->waiting_rcv_msgs_lock);
> >
> > intf->run_to_completion = 1;
> > +restart:
> > intf->handlers->set_run_to_completion(intf->send_info, 1);
> > +
> > + if (intf->curr_msg) {
> > + /*
> > + * This can happen if the kernel panics before
> > + * setting msg to smi_info->waiting_msg or while
> > + * processing a response. For the former case, we
> > + * resend the message by re-queueing it. For the
> > + * latter case, we simply ignore it because handling
> > + * response is not much meaningful in the panic
> > + * context.
> > + */
> > +
> > + /*
> > + * Since we want to send the current message first,
> > + * re-queue it into the high-prioritized queue.
> > + */
> > + if (intf->curr_msg->flags & IPMI_MSG_RESEND_ON_PANIC)
> > + list_add(&intf->curr_msg->link,
> > + &intf->hp_xmit_msgs);
> > +
> > + intf->curr_msg = NULL;
> > + }
> > +
> > + if (!list_empty(&intf->hp_xmit_msgs) ||
> > + !list_empty(&intf->xmit_msgs)) {
> > + /*
> > + * This can happen if the kernel panics while
> > + * processing a response. Kick the queue and restart.
> > + */
> > + smi_recv_tasklet((unsigned long)intf);
> > + goto restart;
> > + }
> > }
> >
> > #ifdef CONFIG_IPMI_PANIC_EVENT
> > diff --git a/drivers/char/ipmi/ipmi_si_intf.c b/drivers/char/ipmi/ipmi_si_intf.c
> > index 814b7b7..c5c7806 100644
> > --- a/drivers/char/ipmi/ipmi_si_intf.c
> > +++ b/drivers/char/ipmi/ipmi_si_intf.c
> > @@ -383,7 +383,6 @@ static enum si_sm_result start_next_msg(struct smi_info *smi_info)
> > int err;
> >
> > smi_info->curr_msg = smi_info->waiting_msg;
> > - smi_info->waiting_msg = NULL;
> > debug_timestamp("Start2");
> > err = atomic_notifier_call_chain(&xaction_notifier_list,
> > 0, smi_info);
> > @@ -401,6 +400,7 @@ static enum si_sm_result start_next_msg(struct smi_info *smi_info)
> > rv = SI_SM_CALL_WITHOUT_DELAY;
> > }
> > out:
> > + smi_info->waiting_msg = NULL;
> > return rv;
> > }
> >
> > @@ -804,6 +804,9 @@ static enum si_sm_result smi_event_handler(struct smi_info *smi_info,
> > {
> > enum si_sm_result si_sm_result;
> >
> > + if (smi_info->curr_msg)
> > + smi_info->curr_msg->flags &= ~(IPMI_MSG_RESEND_ON_PANIC);
> > +
> > restart:
> > /*
> > * There used to be a loop here that waited a little while
> > diff --git a/include/linux/ipmi_smi.h b/include/linux/ipmi_smi.h
> > index ba57fb1..1200872 100644
> > --- a/include/linux/ipmi_smi.h
> > +++ b/include/linux/ipmi_smi.h
> > @@ -47,6 +47,9 @@
> > /* Structure for the low-level drivers. */
> > typedef struct ipmi_smi *ipmi_smi_t;
> >
> > +/* Flags for flags member of struct ipmi_smi_msg */
> > +#define IPMI_MSG_RESEND_ON_PANIC 1 /* If set, resend in panic_event() */
> > +
> > /*
> > * Messages to/from the lower layer. The smi interface will take one
> > * of these to send. After the send has occurred and a response has
> > @@ -75,6 +78,8 @@ struct ipmi_smi_msg {
> > /* Will be called when the system is done with the message
> > (presumably to free it). */
> > void (*done)(struct ipmi_smi_msg *msg);
> > +
> > + int flags;
> > };
> >
> > struct ipmi_smi_handlers {
> >
> >