[email protected], le dim. 28 sept. 2025 20:37:01 +0200, a ecrit:
>
>
>
> Sep 25, 2025, 21:02 by [email protected]:
>
> > Hello,
> >
> > [email protected], le jeu. 25 sept. 2025 13:11:07 +0200, a ecrit:
> >
> >> Hi SamuelI found something in kmsg_putchar.
> >>
> >> if kmsg_write_offset +1 == kmsg_read_offset (mod KMSGBUFFSIZE)
> >> it just discards the character which leads to the previous message being
> >> malformed.
> >>
> >
> > Which is kind of expected: we are overflowing.
> >
> >> Should it in that case also terminate the current message by setting the
> >> previously written char (at kmsg_write_offset -1 mod KMSGBUFFERSIZE) to
> >> \n ?
> >>
> >
> > I'd rather not alter the content.
> >
> >> I tried this patch and it fixes the problem. Some messages are being lost
> >> but I don't know what else to do with them. (i am reusing the offset
> >> variable that is not needed anymore)
> >> I guess I encounter this because at boot time a lot of messages are being
> >> written to the buffer in a short time before the syslogd can start
> >> emptying them so increasing the buffer prevents the early wraparound.
> >>
> >
> > Yes, there is no way around that issue except increasing the buffer size
> > if our boot log is indeed quite verbose (we can indeed do that). Perhaps
> > we'd rather instead drop some messages which are not really that useful.
> >
>
> I am overflowing by just a couple of messages, with the increased buffer I
> can read 4438 bytes.
> > That being said, dropping the latest characters being put might not be
> > the best, I'd rather say drop the oldest characters so that you are sure
> > you have the latest information. And then you'll have '\n'.
> >
> This essentially means writing the char as usual but when after that write
> and read offset are equal incrementing the read offset as well.
>
> Something like this?
Yes, exactly.
Samuel
> --8<---------------cut here---------------start------------->8---
> diff --git a/device/kmsg.c b/device/kmsg.c
> index e5b518e6..bb72930d 100644
> --- a/device/kmsg.c
> +++ b/device/kmsg.c
> @@ -217,7 +217,6 @@ void
> kmsg_putchar (int c)
> {
> io_req_t ior;
> - int offset;
> spl_t s = -1;
>
> /* XXX: cninit is not called before cnputc is used. So call kmsginit
> @@ -230,22 +229,20 @@ kmsg_putchar (int c)
>
> if (spl_init)
> s = simple_lock_irq (&kmsg_lock);
> - offset = kmsg_write_offset + 1;
> - if (offset == KMSGBUFSIZE)
> - offset = 0;
> -
> - if (offset == kmsg_read_offset)
> - {
> - /* Discard C. */
> - if (spl_init)
> - simple_unlock_irq (s, &kmsg_lock);
> - return;
> - }
>
> kmsg_buffer[kmsg_write_offset++] = c;
> if (kmsg_write_offset == KMSGBUFSIZE)
> kmsg_write_offset = 0;
>
> + if(kmsg_write_offset == kmsg_read_offset)
> + {
> + /* Drop first unread char */
> + kmsg_read_offset++;
> + if (kmsg_read_offset == KMSGBUFSIZE)
> + kmsg_read_offset = 0;
> + }
> +
> +
> while ((ior = (io_req_t) dequeue_head (&kmsg_read_queue)) != NULL)
> iodone (ior);
> --8<---------------cut here---------------end--------------->8---
>
> I think I would also prefer this over just dropping the character.
>
>
> > Samuel
> >
>
--
Samuel
/*
* Oops. The kernel tried to access some bad page. We'll have to
* terminate things with extreme prejudice.
*/
die_if_kernel("Oops", regs, error_code);
(From linux/arch/i386/mm/fault.c)
