Re: [PATCH 4/5] IPMI: Add console oops catcher

[Andrew Morton <akpm@linux-foundation.org>]

On Tue, 03 Mar 2009 09:21:23 -0600
Corey Minyard <minyard@acm.org> wrote:

> From: Hiroshi Shimamoto <h-shimamoto@ct.jp.nec.com>
> 
> Console messages on oops or panic are very important to investigate problem.
> Logging oops or panic messages to SEL is useful because SEL is a non
> volatile memory.
> 
> Implement a console driver to log messages to SEL when oops_in_progress is
> not zero. The first message just after oops, panic or every 10 seconds from
> last timestamp are logged as OEM event with timestamp, others are logged as
> OEM event without timestamp.
> 
> Enable config IPMI_OOPS_CONSOLE and add console=ttyIPMI to kernel command
> line to log panic or oops messages to IPMI SEL.
> 
> The number of entries for this output is limited by msg_limit paramter,
> and the default value is 100.
> 
>
> ...
>
> +static void ipmi_oops_console_log_to_sel(int timestamp)
> +{
> +	struct ipmi_system_interface_addr si;
> +	struct kernel_ipmi_msg msg;
> +	unsigned int len;
> +	unsigned char data[16];
> +	unsigned char my_addr;
> +
> +	if (!oops_user || !msg_len || msg_count >= msg_limit)
> +		return;
> +
> +	len = min((timestamp ? SEL_MSGSIZE_TIMESTAMP : SEL_MSGSIZE), msg_len);
> +
> +	si.addr_type = IPMI_SYSTEM_INTERFACE_ADDR_TYPE;
> +	si.channel = IPMI_BMC_CHANNEL;
> +	si.lun = 0;
> +
> +	msg.netfn = IPMI_NETFN_STORAGE_REQUEST;
> +	msg.cmd = IPMI_ADD_SEL_ENTRY_CMD;
> +	msg.data = data;
> +	msg.data_len = 16;
> +
> +	memset(data, 0, sizeof(data));
> +	if (timestamp) {
> +		len = min(SEL_MSGSIZE_TIMESTAMP, msg_len);
> +		data[2] = 0xd1; /* OEM event with timestamp */
> +		memcpy(data + 10, msg_ptr, len);
> +		msg_jiffies = jiffies; /* save jiffies at timestamp */
> +	} else {
> +		len = min(SEL_MSGSIZE, msg_len);
> +		data[2] = 0xf1; /* OEM event without timestamp */
> +		memcpy(data + 3, msg_ptr, len);
> +	}
> +
> +	preempt_disable();

This reader is unable to determine what the mystery preempt_disable()
is here for.  It needs a comment, please.


> +	if (ipmi_get_my_address(oops_user, 0, &my_addr))
> +		goto out;
> +
> +	atomic_set(&oops_counter, 2);

What happens if two CPUs oops at the same time (which is fairly common)?

> +	if (ipmi_request_supply_msgs(oops_user, (struct ipmi_addr *)&si,
> +				     0, &msg, NULL, &oops_smi_msg,
> +				     &oops_recv_msg, 1) < 0)
> +		goto out;
> +	while (atomic_read(&oops_counter) > 0) {
> +		ipmi_poll_interface(oops_user);
> +		cpu_relax();
> +	}

It would be prudent to have a timeout in this loop.  The machine has
crashed and subsystems and hardware and memory are in an unknown and
possibly bad state.

> +	++msg_count;
> +	msg_len -= len;
> +	msg_ptr = msg_len ? &oops_msg[len] : oops_msg;
> +out:
> +	preempt_enable();
> +}
> +
>
> ...
>
> +static struct console oops_console = {
> +	.name	= "ttyIPMI",
> +	.write	= ipmi_oops_console_write,
> +	.unblank = ipmi_oops_console_sync,
> +	.index	= -1,
> +};

This looks like we're abusing the "unblank" method.

If you think that the console subsystem is missing some capability
which this driver needs then please do prefer to fix the console
subsystem, rather than awkwardly making things fit.

> +static void ipmi_oops_recv(struct ipmi_recv_msg *msg, void *data)
> +{
> +	ipmi_free_recv_msg(msg);
> +}
> +
>
> ...
>