Re: [PATCH] depmod: ignore related modules in depmod_report_cycles

[Mian Yousaf Kaukab <yousaf.kaukab@suse.com>]

On Mon, 2016-11-07 at 12:23 -0800, Bjorn Andersson wrote:
> On Mon 07 Nov 09:27 PST 2016, Mian Yousaf Kaukab wrote:
> 
> > 
> > On Sat, 2016-11-05 at 21:50 -0200, Lucas De Marchi wrote:
> > > 
> > > Hi,
> > > 
> > > On Fri, Nov 4, 2016 at 7:33 AM, Mian Yousaf Kaukab
> > > <yousaf.kaukab@suse.com> wrote:
> > > > 
> > > > 
> > > > Only print actual cyclic dependencies. Don't print count of all
> > > > the
> > > > modules as it includes other modules which have dependencies
> > > > but
> > > > not
> > > > necessarily cyclic.
> > > > 
> > > > Printing related modules causes buffer overflow as m->modnamesz 
> > > > is
> > > > not
> > > > included in buffer size calculations (loop == m is never true).
> > > > This buffer overflow causes kmod to crash.
> > > > 
> > > > Reported-by: Andreas Färber <afaerber@suse.de>
> > > > Signed-off-by: Mian Yousaf Kaukab <yousaf.kaukab@suse.com>
> > > > ---
> > > > As count of modules in cyclic dependency chain is not known at
> > > > the
> > > > start of this function, so it is not printed anymore. The
> > > > output
> > > > when cyclic dependency is detected is changed as following:
> > > > 
> > > > Old:
> > > > depmod: ERROR: Found 8 modules in dependency cycles!
> > > > 
> > > > New:
> > > > depmod: ERROR: Modules found in dependency cycles!
> > > 
> > > I think it would be good to fix this problem, but retaining the
> > > behavior.
> > Would it be OK if the modules names are printed first and count is
> > printed afterward. Something like following: 
> > 
> > DEPMOD  4.9.0-rc4-default
> > depmod: ERROR: Cycle detected: qcom_wcnss_iris -> qcom_wcnss ->
> > qcom_wcnss_iris
> > depmod: ERROR: Found 2 modules in dependency cycles!
> > Makefile:1227: recipe for target '_modinst_post' failed
> > 
> > In this way modules names can still be printed in the loop and then
> > the
> > exact count of modules involved in cyclic dependency is printed at
> > the
> > end of the depmod_report_cycles().
> > 
> > > 
> > > 
> > > My first reaction to this is "how do I reproduce the issue?".
> > > This
> > > number so far does tell us how many modules are involved in
> > > loops. 
> > There is more info at the following link:
> > https://bugzilla.opensuse.org/show_bug.cgi?id=1008186
> > 
> > You can reproduce the issue by enabling CONFIG_QCOM_WCNSS_PIL as
> > module
> > in v4.9-rc4 for arm64. 
> > +CONFIG_REMOTEPROC=m
> > +CONFIG_QCOM_MDT_LOADER=m
> > -# CONFIG_QCOM_WCNSS_PIL is not set
> > +CONFIG_QCOM_WCNSS_IRIS=m
> > +CONFIG_QCOM_WCNSS_PIL=m
> > 
> 
> I added some debugging prints to track the stack management in
> depmod_report_cycles().
> 
> remoteproc, mdt_loader, wcnss_iris and wcnss_pil are all roots and we
> have dependencies like this.
> 
>        /---(rproc)
>        |     ^
>        |     |
> (pil) -+-> (mdt)
>   ^    |
>   |    v
>   +- (iris)
AFAICT this is not correct. Dependencies are like following:

pil -> rproc
pil -> mtd -> rproc
pil -> iris -> pil

> 
> 1) We pick rproc as first root, mark that as visited and see that
> we're
>    done.
> 
> 2) We pick mdt_loader as second root, we push remoteproc to the stack
>    and find that it's already visited, so we found a loop!
> 
>      mdt_loader -> remoteproc
This is not cyclic.

> 
> 3) We pick iris as third root, we push wcnss which pushes remoteproc,
>    mdt_loader and iris. All three have already been visited from root
> #1
>    and #2, so we find that there's a loop in:
> 
>      iris -> wcnss -> iris
>      iris -> wcnss -> mdt
>      iris -> wcnss -> remoteproc
> 
> Only one of these cases are actually a cycle, but as we don't reset
> visited between the searches we can't tell. Further more, if there
> was a
> dependency from iris -> remoteproc that would have shown up earlier
> and
> marked iris->visited and when we get to step #3 we would just have
> bailed directly - completely missing the cycle.
In case we have more than one cyclic dependencies around same modules?
May be we should add a testcase for such a scenario so that its easy to
understand and reproduce.

> 
> So I think we need to reset the visited list on each run of the DFS
> from
> each root.
> 
> Regards,
> Bjorn

BR,
Yousaf