lkml.org 
[lkml]   [2017]   [Oct]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH v2 5/5] of/fdt: only store the device node basename in full_name
From
Date
On Wed, 2017-10-18 at 10:44 -0500, Rob Herring wrote:
> On Wed, Oct 18, 2017 at 10:12 AM, Alan Tull <atull@kernel.org> wrote:
> > On Tue, Oct 17, 2017 at 6:51 PM, Frank Rowand <frowand.list@gmail.com> wrote:
> >> On 10/17/17 14:46, Rob Herring wrote:
> >>> On Tue, Oct 17, 2017 at 4:32 PM, Alan Tull <atull@kernel.org> wrote:
> >>>> On Mon, Aug 21, 2017 at 10:16 AM, Rob Herring <robh@kernel.org> wrote:
> >>>>
> >>>> Hi Rob,
> >>>>
> >>>>> With dependencies on a statically allocated full path name converted to
> >>>>> use %pOF format specifier, we can store just the basename of node, and
> >>>>> the unflattening of the FDT can be simplified.
> >>>>>
> >>>>> This commit will affect the remaining users of full_name. After
> >>>>> analyzing these users, the remaining cases should only change some print
> >>>>> messages. The main users of full_name are providing a name for struct
> >>>>> resource. The resource names shouldn't be important other than providing
> >>>>> /proc/iomem names.
> >>>>>
> >>>>> We no longer distinguish between pre and post 0x10 dtb formats as either
> >>>>> a full path or basename will work. However, less than 0x10 formats have
> >>>>> been broken since the conversion to use libfdt (and no one has cared).
> >>>>> The conversion of the unflattening code to be non-recursive also broke
> >>>>> pre 0x10 formats as the populate_node function would return 0 in that
> >>>>> case.
> >>>>>
> >>>>> Signed-off-by: Rob Herring <robh@kernel.org>
> >>>>> ---
> >>>>> v2:
> >>>>> - rebase to linux-next
> >>>>>
> >>>>> drivers/of/fdt.c | 69 +++++++++-----------------------------------------------
> >>>>> 1 file changed, 11 insertions(+), 58 deletions(-)
> >>>>
> >>>> I've just updated to the latest next branch and am finding problems
> >>>> applying overlays. Reverting this commit alleviates things. The
> >>>> errors I get are:
> >>>>
> >>>> [ 88.498704] OF: overlay: Failed to apply prop @/__symbols__/clk_0
> >>>> [ 88.513447] OF: overlay: apply failed '/__symbols__'
> >>>> [ 88.518423] create_overlay: Failed to create overlay (err=-12)
> >>>
> >>> Frank's series with overlay updates should fix this.
> >>
> >> Yes, it does:
> >>
> >> [PATCH v3 11/12] of: overlay: remove a dependency on device node full_name
> >
> > Thanks for the fast response. I fetched the dt/next branch to test
> > this but there are sufficient changes that Pantelis' "OF: DT-Overlay
> > configfs interface (v7)" is broken now. I've been adding that
> > downstream since 4.4. We're using it as an interface for applying
> > overlays to program FPGAs. If we fix it again, is there any chance
> > that can go upstream now?
>
> With a drive-by posting once every few years, no.
>

I take offense to that. There's nothing changed in the patch for years.
Reposting the same patch without changes would achieve nothing.

> The issue remains that the kernel is not really setup to deal with any
> random property or node to be changed at any point in run-time. I
> think there needs to be some restrictions around what the overlays can
> touch. We can't have it be wide open and then lock things down later
> and break users. One example of what you could do is you can only add
> sub-trees to whitelisted nodes. That's probably acceptable for your
> usecase.
>

Defining what can and what cannot be changed is not as trivial as a
list of white-listed nodes.

In some cases there is a whole node hierarchy being inserted (like in
a FPGA). In others, it's merely changing a status property to "okay" and
a few device parameters.

The real issue is that the kernel has no way to verify that a given
device tree, either at boot time or at overlay application time, is
correct.

When the tree is wrong at boot-time you'll hang (if you're lucky).
If the tree is wrong at run-time you'll get some into some unidentified
funky state.

Finally what is, and what is not 'correct' is not for the kernel to
decide arbitrarily, it's a matter of policy, different for each
use-case.

> Rob

Regards

-- Pantelis


\
 
 \ /
  Last update: 2017-10-22 17:15    [W:0.094 / U:1.068 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site