Message-ID: <AM6PR04MB3976C893BE91E755D439EDFFECCB0@AM6PR04MB3976.eurprd04.prod.outlook.com>
Date:   Thu, 10 Dec 2020 09:05:09 +0000
From:   Madalin Bucur <madalin.bucur@....com>
To:     Patrick Havelange <patrick.havelange@...ensium.com>,
        "David S. Miller" <davem@...emloft.net>,
        Jakub Kicinski <kuba@...nel.org>,
        "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>
Subject: RE: [PATCH net 1/4] net: freescale/fman: Split the main resource
 region reservation

> -----Original Message-----
> From: Patrick Havelange <patrick.havelange@...ensium.com>
> Sent: 10 December 2020 10:49
> To: Madalin Bucur <madalin.bucur@....com>; David S. Miller
> <davem@...emloft.net>; Jakub Kicinski <kuba@...nel.org>;
> netdev@...r.kernel.org; linux-kernel@...r.kernel.org
> Subject: Re: [PATCH net 1/4] net: freescale/fman: Split the main resource
> region reservation
> 
> On 2020-12-09 19:55, Madalin Bucur wrote:
> >> -----Original Message-----
> >> From: Patrick Havelange <patrick.havelange@...ensium.com>
> >> Sent: 09 December 2020 16:17
> >> To: Madalin Bucur <madalin.bucur@....com>; David S. Miller
> >> <davem@...emloft.net>; Jakub Kicinski <kuba@...nel.org>;
> >> netdev@...r.kernel.org; linux-kernel@...r.kernel.org
> >> Subject: Re: [PATCH net 1/4] net: freescale/fman: Split the main
> resource
> >> region reservation
> >>
> >>>>> area. I'm assuming this is the problem you are trying to address
> here,
> >>>>> besides the stack corruption issue.
> >>>>
> >>>> Yes exactly.
> >>>> I did not add this behaviour (having a main region and subdrivers
> using
> >>>> subregions), I'm just trying to correct what is already there.
> >>>> For example: this is some content of /proc/iomem for one board I'm
> >>>> working with, with the current existing code:
> >>>> ffe400000-ffe4fdfff : fman
> >>>>      ffe4e0000-ffe4e0fff : mac
> >>>>      ffe4e2000-ffe4e2fff : mac
> >>>>      ffe4e4000-ffe4e4fff : mac
> >>>>      ffe4e6000-ffe4e6fff : mac
> >>>>      ffe4e8000-ffe4e8fff : mac
> >>>>
> >>>> and now with my patches:
> >>>> ffe400000-ffe4fdfff : /soc@...000000/fman@...000
> >>>>      ffe400000-ffe480fff : fman
> >>>>      ffe488000-ffe488fff : fman-port
> >>>>      ffe489000-ffe489fff : fman-port
> >>>>      ffe48a000-ffe48afff : fman-port
> >>>>      ffe48b000-ffe48bfff : fman-port
> >>>>      ffe48c000-ffe48cfff : fman-port
> >>>>      ffe4a8000-ffe4a8fff : fman-port
> >>>>      ffe4a9000-ffe4a9fff : fman-port
> >>>>      ffe4aa000-ffe4aafff : fman-port
> >>>>      ffe4ab000-ffe4abfff : fman-port
> >>>>      ffe4ac000-ffe4acfff : fman-port
> >>>>      ffe4c0000-ffe4dffff : fman
> >>>>      ffe4e0000-ffe4e0fff : mac
> >>>>      ffe4e2000-ffe4e2fff : mac
> >>>>      ffe4e4000-ffe4e4fff : mac
> >>>>      ffe4e6000-ffe4e6fff : mac
> >>>>      ffe4e8000-ffe4e8fff : mac
> >>>>
> >>>>> While for the latter I think we can
> >>>>> put together a quick fix, for the former I'd like to take a bit of
> >> time
> >>>>> to select the best fix, if one is really needed. So, please, let's
> >> split
> >>>>> the two problems and first address the incorrect stack memory use.
> >>>>
> >>>> I have no idea how you can fix it without a (more correct this time)
> >>>> dummy region passed as parameter (and you don't want to use the first
> >>>> patch). But then it will be useless to do the call anyway, as it
> won't
> >>>> do any proper verification at all, so it could also be removed
> entirely,
> >>>> which begs the question, why do it at all in the first place (the
> >>>> devm_request_mem_region).
> >>>>
> >>>> I'm not an expert in that part of the code so feel free to correct me
> >> if
> >>>> I missed something.
> >>>>
> >>>> BR,
> >>>>
> >>>> Patrick H.
> >>>
> >>> Hi, Patrick,
> >>>
> >>> the DPAA entities are described in the device tree. Adding some
> >> hardcoding in
> >>> the driver is not really the solution for this problem. And I'm not
> sure
> >> we have
> >>
> >> I'm not seeing any problem here, the offsets used by the fman driver
> >> were already there, I just reorganized them in 2 blocks.
> >>
> >>> a clear problem statement to start with. Can you help me on that part?
> >>
> >> - The current call to __devm_request_region in fman_port.c is not
> correct.
> >>
> >> One way to fix this is to use devm_request_mem_region, however this
> >> requires that the main fman would not be reserving the whole region.
> >> This leads to the second problem:
> >> - Make sure the main fman driver is not reserving the whole region.
> >>
> >> Is that clearer like this ?
> >>
> >> Patrick H.
> 
> Hi,

Hi, Patrick,

> > The overlapping IO areas result from the device tree description, that
> in turn
> > mimics the HW description in the manual. If we really want to remove the
> nesting,
> > we should change the device trees, not the drivers.
> 
> But then that change would not be compatible with the existing device
> trees in already existing hardware. I'm not sure how to handle that case
> properly.

One needs to be backwards compatible with the old device trees, so we do not
really have a simple answer, I know.

> > If we want to hack it,
> > instead of splitting ioremaps, we can reserve 4 kB in the FMan driver,
> > and keep the ioremap as it is now, with the benefit of less code churn.
> 
> but then the ioremap and the memory reservation do not match. Why bother
> at all then with the mem reservation, just ioremap only and be done with
> it. What I'm saying is, I don't see the point of having a "fake"
> reservation call if it does not correspond that what is being used.

The reservation is not fake, it just covering the first portion of the ioremap.
Another hypothetical FMan driver would presumably reserve and ioremap starting
from the same point, thus the desired error would be met.

Regarding removing reservation altogether, yes, we can do that, in the child
device drivers. That will fix that use after free issue you've found and align
with the custom, hierarchical structure of the FMan devices/drivers. But would
leave them without the double use guard we have when using the reservation.

> > In the end, what the reservation is trying to achieve is to make sure
> there
> > is a single driver controlling a certain peripeheral, and this basic
> > requirement would be addressed by that change plus devm_of_iomap() for
> child
> > devices (ports, MACs).
> 
> Again, correct me if I'm wrong, but with the fake mem reservation, it
> would *not* make sure that a single driver is controlling a certain
> peripheral.

Actually, it would. If the current FMan driver reserves the first part of the FMan
memory, then another FMan driver (I do not expect a random driver trying to map the
FMan register area) would likely try to use that same part (with the same or
a different size, it does not matter, there will be an error anyway) and the
reservation attempt will fail. If we fix the child device drivers, then they
will have normal mappings and reservations.

> My point is, either have a *correct* mem reservation, or don't have one
> at all. There is no point in trying to cheat the system.

Now we do not have correct reservations for the child devices because the
parent takes it all for himself. Reduce the parent reservation and make room
for correct reservations for the child. The two-sections change you've made may
try to be correct but it's overkill for the purpose of detecting double use.
And I also find the patch to obfuscate the already not so readable code so I'd
opt for a simpler fix.

Madalin

Powered by blists - more mailing lists