On Wed, Oct 28, 2020 at 03:26:01PM +1100, David Gibson wrote:
> On Tue, Oct 27, 2020 at 02:55:17PM -0500, Rob Herring wrote:
> > On Tue, Oct 27, 2020 at 10:58 AM André Przywara <andre.przywara@xxxxxxx> wrote:
> > >
> > > On 26/10/2020 21:51, Rob Herring wrote:
> > > > On Thu, Oct 22, 2020 at 10:23 AM Tom Rini <trini@xxxxxxxxxxxx> wrote:
> > > >> On Fri, Oct 23, 2020 at 01:58:04AM +1100, David Gibson wrote:
> > > >>> On Thu, Oct 22, 2020 at 08:32:54AM -0400, Tom Rini wrote:
> > > >>>> On Thu, Oct 22, 2020 at 03:00:13PM +1100, David Gibson wrote:
> > > >>>>> On Wed, Oct 21, 2020 at 06:49:14PM -0400, Tom Rini wrote:
> > > >
> > > > [...]
> > > >
> > > >>>>>> But what does all of this _mean_ ? I kinda think I have an answer now.
> > > >>>>>> One of the things that sticks out is 6dcb8ba408ec adds a lot and
> > > >>>>>> 11738cf01f15 reduces it just a little.
> > > >>>>>
> > > >>>>> Ah, that's a tricky one. If we don't handle unaligned accesses we
> > > >>>>> instead get intermittent bug reports where it just crashes.
> > > >>>>
> > > >>>> We really need to talk about that then. There was a problem of people
> > > >>>> turning off the sanity check for making sure the entire device tree was
> > > >>>> aligned and then having everything crash.
> > > >>>
> > > >>> Ok... I'm not really sure where you're going with that thought.
> > > >>
> > > >> In my reading of the mailing list history of how this issue came up,
> > > >> it was someone was booting a dragonboard or something, and they (or
> > > >> rather, the board maintainer set by default) the flag to use the device
> > > >> tree wherever it is in memory and NOT to relocate it to a properly
> > > >> aligned address. This in turn lead to the kernel getting an unaligned
> > > >> device tree and everything crashing. The "I know what I'm doing" flag
> > > >> was set, violated the documented requirements for device trees need to
> > > >> reside in memory and everything blew up.
> > > >>
> > > >> After that it was noticed that there could be some internal
> > > >> mis-alignment and if you tried those accesses on a CPU that doesn't
> > > >> support doing those reads easily there could be problems, but that's not
> > > >> a common at all case (as noted by it not having been seen in practice).
> > > >
> > > > Nor a problem on many environments to begin with. More below...
> > > >
> > > >>>>> I suppose we could add an ASSUME_ALIGNED_ACCESS flag, and it will just
> > > >>>>> break for either an unaligned dtb (unlikely) or if you attempt to load
> > > >>>>> an unaligned value from a property (more likely, but don't add the
> > > >>>>> flag if you're not sure you don't need it).
> > > >>>>
> > > >>>> So long as it's abstracted in such a way that we don't grow the size of
> > > >>>> everything again, yes, that is the right way forward I think.
> > > >>>
> > > >>> All the ASSUME flags should be resolved at compile time (at least with
> > > >>> normal optimization levels enabled in the compiler), so testing for
> > > >>> those shouldn't increase size at all. If they do, something is wrong.
> > > >>
> > > >> I'm saying that how ever this new ASSUME flag is done, it needs to be
> > > >> done in such a way the compiler really will be smart about it. So
> > > >> something like making a new function that does fdt64_ld() if we aren't
> > > >> ASSUME_ALIGNED_ACCESS and fdt64_to_cpu() if we are
> > > >> ASSUME_ALIGNED_ACCESS.
> > > >
> > > > Ah, unaligned accesses again... To summarize, both performance and
> > > > size suffer with not doing unaligned accesses.
> > > >
> > > > Why not a HAS_UNALIGNED_ACCESS flag instead (or the inverse) that will
> > > > do unaligned accesses? That would be more aligned with what the system
> > > > can support rather than sanity checking associated with ASSUME_*.
>
> So, there are kind of two things here, (1) is "my platform can handle
> unaligned accesses" and (2) is "assume I don't need unaligned
> accesses". We can use the fast & small versions of fdt32_ld() etc. if
> either is true. However we need to consider those separately, because
> they can be independently true (or not) for different reasons. (1)
> depends on the hardware, whereas (2) depends on how you're using dtc,
> and, see below, you may need at least unaligned-handling fdt64_ld() in
> more cases than you think.
>
> > > > To repeat from last time, everything ARMv6 and up can do unaligned
> > > > accesses if enabled.
> > >
> > > But that requires the MMU to be enabled, doesn't it? If I read the ARM
> > > ARM correctly, unaligned accesses always trap on device memory,
> > > regardless of SCTLR.A. And without the MMU enabled everything is device
> > > memory. We compile U-Boot with -mno-unaligned-access/-mstrict-align to
> > > cope with that, and that most likely affects libfdt as well?
> >
> > Ah yes, I think you are right.
> >
> > In that case, seems like we should figure out whether (internal)
> > unaligned accesses are possible with dtc generated dtbs at least
> > rather than just "not a common at all case (as noted by it not having
> > been seen in practice)." I'm sure David will point out that not all
> > dtbs come from dtc, but all the ones u-boot deals with do in
> > reality.
>
> Assuming the blob itself is 8-byte aligned in memory, then all
> structural elements (i.e. the tree metadata) of a compliant dtb will
> be naturally aligned. The spec requires 8-byte alignment of the mem
> reserve block w.r.t. the base of the blob and 4 byte aligned structure
> block w.r.t. the base of the blob. Likewise the layout of the mem
> reserve block will preserve 8-byte alignment of all the 64-bit values
> it contains, assuming the block itself starts 8-byte aligned.
> Similarly the structure blob will preserve 4-byte alignment of all its
> tags and other structural data (this amounts to requiring an alignment
> gap after node names and property values).
>
> However, "all structural elements" does not include values within
> property values themselves. Assuming propery alignment of the blocks
> and the blob itself, then all property values will *begin* 4 byte
> aligned. However that leaves two relevant cases:
>
> a) 64-bit property values may be 4-byte aligned but not 8-byte
> aligned
> b) complex property values including both strings and integers
> typically use a packed representation with no alignment gaps.
> Such property structures are usually avoided in modern bindings,
> but they definitely exist in a bunch of older bindings. Obviously
> that means that integer values sitting after arbitrary length
> strings may not have any natural alignment
>
> So acccesses made by libfdt internally should be safe(*) assuming the
> blob itself is loaded 8-byte aligned, and the dtb is compliant.
> However the libfdt user may hit both problems (a) and (b) getting
> things they actually want from the tree. fdt{32,64}_{ld,st}() are
> intended to handle those cases, so that they're useful for the caller
> to pull things from properties as well as for libfdt internal
> accesses.
>
> (*) There are a number of other functions that looked like they might
> be dangerous for case (a) because they are based on 64-bit
> property values: fdt_setprop_inplace_u64(), fdt_property_u64(),
> fdt_setprop_u64(), fdt_appendprop_u64() and
> fdt_appendprop_addrrange(). However I think they're actually
> ok, because the way they're built in terms of other functions
> means there's implicitly a memcpy() from a byte buffer.
>
> > > Also some 32-bit ARM platforms run U-Boot proper with the MMU disabled
> > > all the time, and I know of at least the sunxi-aarch64 SPL running with
> > > the MMU off as well.
> >
> > I'm making a mental note of this for the next time performance issues come up.
>
> Right, running early with MMU off is definitely a real use case for
> libfdt. For similar reasons we can't assume we have an OS which will
> trap and handle unaligned accesses, which we might for a more
> conventional userspace library.
>
> This kind of underscores why I'm a bit hesitant to introduce "my
> platform handles unaligned acccesses" flag. Not only does it require
> detailed knowledge of the target CPU, but it can also depend on
> exactly what mode that hardware is in.
Can you please note the existing user(s) where we have just the right
combination of factors and so everything fails?
--
Tom
Attachment:
signature.asc
Description: PGP signature
