On Fri, Jul 10, 2020 at 02:19:23PM +0200, Petr Mladek wrote:
> On Thu 2020-07-09 12:54:15, Bruno Meneguele wrote:
> > The commit 625d3449788f ("Revert "kernel/printk: add kmsg SEEK_CUR
> > handling"") reverted a change done to the return value in case a SEEK_CUR
> > operation was performed for kmsg buffer based on the fact that different
> > userspace apps were handling the new return value (-ESPIPE) in different
> > ways, breaking them.
> >
> > At the same time -ESPIPE was the wrong decision because kmsg /does support/
> > seek() but doesn't follow the "normal" behavior userspace is used to.
> > Because of that and also considering the time -EINVAL has been used, it was
> > decided to keep this way to avoid more userspace breakage.
> >
> > This patch adds an official statement to the kmsg documentation pointing to
> > the current return value for SEEK_CUR, -EINVAL, thus userspace libraries and
> > apps can refer to it for a definitive guide on what to expected.
> >
> > --- a/Documentation/ABI/testing/dev-kmsg
> > +++ b/Documentation/ABI/testing/dev-kmsg
> > @@ -56,6 +56,11 @@ Description: The /dev/kmsg character device node
> > provides userspace access
> > seek after the last record available at the time
> > the last SYSLOG_ACTION_CLEAR was issued.
> >
> > + Considering that the seek operation is supported but has
> > + special meaning to the device, any attempt to seek specific
> > + positions on the buffer (i.e. using SEEK_CUR) results in an
> > + -EINVAL error returned to userspace.
>
> Sigh, I see that devkmsg_llseek() returns -ESPIPE when offset is not
> zero. This is a real mess.
>
> I wonder if we could afford to switch this one to -EINVAL and reduce
> the mess.
> That's a really good question for which I think the answer is something close to: "that's impossible to predict". I could try some in-house (at redhat) automated tests to see if something break, but I doubt it is enough to catch any breakage. > Anyway, for a random reader, it might be pretty unclear what is > exactly special about /dev/kmsg. I wonder if the following might > be more explanatory and strightforward: > > Other seek operations or offsets are not supported because of > the special behavior. The device allows to read or write > only whole variable lenght messages that are stored in > a ring buffer. > > Because of the non-standard behavior also the error values > are non-standardand. -ESPIPE is returned for non-zero offset. > -EINVAL is returned for other operations, e.g. SEEK_CUR. > This behavior is historical and could not be modified > wihtout the risk of breaking userspace. > Yes, no doubt it's better! > > Finally, only few people read documentation. We should add some > warning also to the code. I think about a something like: > > /* > * Be careful when modifying this function!!! > * > * Only few operations are supported because the device works only with > * the entire variable length messages. Non-standard error values are > * returned in the other cases. User space applications might depend > * on this behavior. > */ > Agreed. I'm going to prepare a v2 of the patch adding the comments. Thanks Petr! -- bmeneg PGP Key: http://bmeneg.com/pubkey.txt
signature.asc
Description: PGP signature
