Re: [PATCH 2/2] printk: use the lockless ringbuffer

[John Ogness]

On 2020-02-14, Sergey Senozhatsky <sergey.senozhatsky.work@xxxxxxxxx> wrote:
>>> cat /dev/kmsg
>>> cat: /dev/kmsg: Broken pipe
>>
>> In mainline you can have this "problem" as well. Once the ringbuffer
>> has wrapped, any read to a newly opened /dev/kmsg when a new message
>> arrived will result in an EPIPE. This happens quite easily once the
>> ringbuffer has wrapped because each new message is overwriting the
>> oldest message.
>
> Hmm. Something doesn't add up.
>
> Looking at the numbers, both r->info->seq and prb_first_seq(prb)
> do increase, so there are new messages in the ring buffer
>
>                            u->seq    r->seq    prb_first_seq
> [..]
> cat: devkmsg_read() error 1981080   1982633   1981080
> cat: devkmsg_read() error 1981080   1982633   1981080
> cat: devkmsg_read() error 1981095   1982652   1981095
> cat: devkmsg_read() error 1981095   1982652   1981095
> cat: devkmsg_read() error 1981095   1982652   1981095
> [..]
>
> but 'cat' still wouldn't read anything from the logbuf - EPIPE.
>
> NOTE: I don't run 'cat /dev/kmsg' during the test. I run the test
> first, then I run 'cat /dev/kmsg', after the test, when
> printk-pressure is gone.

Sure. The problem is not the printk-pressure. The problem is you have
data-less records in your ringbuffer (from your previous
printk-pressure). If you used your own program that continued to read
after EPIPE, then you would see the sequence numbers jumping over the
data-less records.

> I can't reproduce it with current logbuf. 'cat' reads from /dev/kmsg
> after heavy printk-pressure test. So chances are some loggers can also
> experience problems. This might be a regression.

Mainline doesn't have data-less records. In mainline such failed
printk's are silently ignored (after attepting truncation). So for
mainline you can only reproduce the overflow case.

1. Boot 5.6.0-rc1 (without any console= slowing down printk)

2. Fill the ringbuffer and let it overflow with:

   $ while true; do echo filling buffer > /dev/kmsg; done &

3. Once you can see the ringbuffer has overflowed (and continues to
   overflow), try to read from /dev/kmsg

   $ strace head /dev/kmsg

In most cases you will see:

read(3, 0x7f7307ac1000, 4096)           = -1 EPIPE (Broken pipe)

Current readers need to be able to handle EPIPE. cat(1) does not and so
(unfortunately) is not a good candidate for reading the ringbuffer.

>>> ...
>>> systemd-journal: devkmsg_read() error 1979281 1982465 1980933
>>> systemd-journal: corrected seq 1982465 1982465
>>> cat: devkmsg_read() error 1980987 1982531 1980987
>>> cat: corrected seq 1982531 1982531
>>> cat: devkmsg_read() error 1981015 1982563 1981015
>>> cat: corrected seq 1982563 1982563
>>
>> The situation with a data-less record is the same as when the ringbuffer
>> wraps: cat is hitting that EPIPE. But re-opening the file descriptor is
>> not going to help because it will not be able to get past that data-less
>> record.
>
> So maybe this is the case with broken 'cat' on my system?

I think it is appropriate for an application to close the descriptor
after an EPIPE. /dev/kmsg is special because the reader should continue
reading anyway.

>> We could implement it such that devkmsg_read() will skip over data-less
>> records instead of issuing an EPIPE. (That is what dmesg does.) But then
>> do we need EPIPE at all? The reader can see that is has missed records
>> by tracking the sequence number, so could we just get rid of EPIPE? Then
>> cat(1) would be a great tool to view the raw ringbuffer. Please share
>> your thoughts on this.
>
> Looking at systemd/src/journal/journald-kmsg.c :
> server_read_dev_kmsg() -EPIPE is just one of the erronos they handle,
> nothing special.

Yes, but what does systemd-journald do when the EPIPE is _not_ returned
and instead there is a jump in the sequence number? Looking at
dev_kmsg_record(), systemd actually does it the way I would hope. It
tracks the sequence number correctly.

    /* Did we lose any? */
    if (serial > *s->kernel_seqnum)
         server_driver_message(s, 0,
                               "MESSAGE_ID="
                               SD_MESSAGE_JOURNAL_MISSED_STR,
                               LOG_MESSAGE("Missed %"PRIu64" kernel messages",
                               serial - *s->kernel_seqnum),
                               NULL);

> Could it be the case that some other loggers would have special
> handling for EPIPE?  I'm not sure, let's look around.
>
> I'd say that EPIPE removal looks OK to me. But before we do that, I'm
> not sure that we have clear understanding of 'cat /dev/kmsg' behaviour
> change.

In mainline, with regard to /dev/kmsg, sequence numbers will never
jump. If there would be a jump (due to lost messages) then EPIPE is
issued. The reader can either:

1. continue reading and see the jump

2. reopen the file descriptor, possibly having missed a ton more
   messages due to reopening, and then start from the oldest available
   message

With my series, #2 is no longer an option because the lost messages
could exist in a part of the ringbuffer not yet overwritten.

If we remove EPIPE, then readers will need to track the sequence number
to identify jumps. systemd-journald does this already. And tools like
cat(1) would "just work" because cat does not care if messages were
lost.

John Ogness