[lkml]   [2018]   [Nov]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH v6 1/3] printk: Add line-buffered printk() API.
On Thu 2018-11-08 11:21:38, Sergey Senozhatsky wrote:
> On (11/07/18 11:21), Petr Mladek wrote:
> What is the problem:
> - we have tons of CPUs, with tons of tasks running on them, with preemption,
> and interrupts, and potentially printk-s coming from various
> contexts/CPUs/tasks etc. so one 'cont' buffer is not enough.
> What is the proposed solution:
> - if 1 is not enough then 16 will do. And if 16 is not enough then this
> is not our problem anymore, it's a kernel misconfiguration and users'
> fault.

I believe that I mentioned this more times. 16 buffers is the first
attempt. We could improve it later in several ways:

+ add more buffers
+ combine buffers on stack, dynamically allocated and static ones.

BTW: I do not remember seeing mixed lines from anything even close to 16
CPUs. Maybe I was just lucky or my memory is leaky.

> Let's have one more look at what we will fix and what we will break.
> 'cont' has premature flushes.
> Why is it good.
> It preserves the correct order of events.
> pr_cont("calling foo->init()....");
> foo->init()
> printk("Can't allocate buffer\n"); // premature flush
> pr_cont("...blah\h");
> Will end up in the logbuf as:
> [12345.123] calling foo->init()....
> [12345.124] Can't allocate buffer
> [12345.125] ...blah
> Where buffered printk will endup as:
> [12345.123] Can't allocate buffer
> [12345.124] calling foo->init().......blah

We will always have this problem with API using explicit buffers.
What do you suggest instead, please?

I am afraid that we are running in cycles. The other serious
alternative was having per-process and per-context buffers
but it was rejected several times.

> Not to mention that buffered printk does not flush on panic.
> So, frankly, as of now, I don't see buffered printk as a 'cont'
> replacement.

The static array of buffers can be flushed on panic.

> If our problem is OOM and lockdep print outs, then let's address only
> those two; let's not "fix" the rest of the kernel, especially the early
> boot, - we can break more things than we can mend.

Do you have any alternative proposal how to handle OOM and lockdep, please?

> [..]
> > I opened this problem once and it got lost. So I did not want to
> > complicate it at this moment.
> - I don't exactly like the completely of the vprintk_buffered. If
> buffered printk is for single line, then it must be for single
> line only.

My undestanding is that the new API is similar to the current cont
buffer from this point of view:

+ buffer size is LOG_LINE_MAX
+ it is flushed when full

The only difference is that it is flushed also when there is a
complete line. Is this a problem, please?

> And I'm not buying the "we will need this for printk origin
> info injection" argument.

I was against this idea several times. The current API does
not do anything like this.

> - It seems that buffered printk attempts to solve too many problems.
> I'd prefer it to address just one.

This API tries to handle continuous lines more reliably.
Do I miss anything, please?

Best Regards,

 \ /
  Last update: 2018-11-08 12:25    [W:0.117 / U:7.724 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site