On Thu, Feb 2, 2023 at 1:56 PM Kent Mcleod <[email protected]> wrote:

> On Fri, Feb 3, 2023 at 8:26 AM Sam Leffler via Devel <[email protected]>
> wrote:
> >
> > I have a target platform with only 4M of memory. When the system image is
> > generated and the shoehorn helper script is used to find a place in
> memory
> > to load the build artifacts it tacks on an extra 4M of memory use (aka
> > fudge_factor). The comment in the code
> > <
> https://github.com/AmbiML/sparrow-seL4_tools/blame/master/cmake-tool/helpers/shoehorn.py#L209
> >
> > says this is to accommodate sel4test_driver. Needless to say this breaks
> on
> > my 4M target platform. So I made the fudge-factor settable from the cmd
> > line with a default of 0 and changed the sel4test build glue to set 4M
> when
> > building elfloader. Works fine for my target platform. But this change
> > breaks building a bootable image for rpi3 (AARCH64=1 bcm28367)--shoehorn
> > places elfloader s.t. it overlaps the image; e.g.
> >
> > ELF-loader started on CPU: ARM Ltd. Cortex-A53 r0p4
> > >   paddr=[335000..51a0ff]
> > > No DTB passed in from boot loader.
> > > Looking for DTB in CPIO archive...found at 378778.
> > > Loaded DTB from 378778.
> > >    paddr=[237000..23afff]
> > > ELF-loading image 'kernel' to 0
> > >   paddr=[0..236fff]
> > >   vaddr=[ffffff8000000000..ffffff8000236fff]
> > >   virt_entry=ffffff8000000000
> > > ELF-loading image 'capdl-loader' to 23b000
> > >   paddr=[23b000..33bfff]
> > >   vaddr=[400000..500fff]
> > >   virt_entry=4009a8
> > > ERROR: image load address overlaps with ELF-loader!
> > > ERROR: Physical address range invalid
> > > ERROR: Could not load user image ELF
> >
> >
> > Debug output of shoehorn for this case:
> >
> > shoehorn: debug: found CPIO identifying sequence b'070701' at offset 0x40
> > > in
> > >
> /usr/local/google/home/sleffler/shodan/out/cantrip/aarch64-unknown-elf/release/elfloader/archive.o
> > > shoehorn: debug: encountered CPIO entry name: kernel.elf
> > > shoehorn: debug: encountered CPIO entry name: kernel.dtb
> > > shoehorn: debug: encountered CPIO entry name: capdl-loader
> > > shoehorn: debug: setting marker to 0x0 (region 0 start)
> > > shoehorn: debug: setting marker to 0x237000 (kernel_end)
> > > shoehorn: debug: setting marker to 0x23b000 (dtb_end)
> > > shoehorn: debug: setting marker to 0x335000 (end of rootserver)
> >
> >
> > So two questions:
> > 1. Where is the 4M under-count of sel4test_driver? (the code indicates
> this
> > might be explained in JIRA SELFOUR-2335 but I couldn't locate it)
>
> Here is the referred to Jira issue, but it doesn't provide any
> additional context: https://sel4.atlassian.net/browse/SELFOUR-2335
>
> shoehorn is attempting to calculate how the kernel and root server
> binaries will be unpacked into memory in order to place the
> elfloader's start address above the unpacked region. shoehorn
> calculates the region by iterating over the PT_LOAD segments from each
> ELF file. The elfloader then unpacks each ELF file at runtime by
> iterating over the PT_LOAD segments.
>
> For some reason, the two implementations don't agree. In your case,
> the offline calculation expects that the root server is loaded from
> [0x23b000, 0x335000) whereas the online calculation attempts:
> [0x23b000, 0x33bfff). Are you able to print the segment headers for
> the root server image you are loading?
>
> I'm guessing (from quickly looking at the code) the issue is that the
> shoehorn calculation only sums the p_memsz amounts for each PT_LOAD
> segment and isn't taking into account any gaps between segments in the
> virtual address space.
>

Yes, that appears to be the issue. readelf of capdl-loader shows:

Program Headers:
  Type           Offset             VirtAddr           PhysAddr
                 FileSiz            MemSiz              Flags  Align
  LOAD           0x0000000000000000 0x0000000000400000 0x0000000000400000
                 0x00000000000a9130 0x00000000000a9130  RWE    0x1000
  LOAD           0x0000000000000000 0x00000000004b0000 0x00000000004b0000
                 0x0000000000000000 0x0000000000050168  RW     0x1000

so there's a gap between the two load segments that isn't accounted for.
Attached is a change that seems to DTRT. It also appears to eliminate the
need for fudge_factor (in quick testing). You'll probably want to write
your own fix as my python fu is basic.

>
> A fudge-factor wouldn't be needed if these two calculations weren't out of
> sync.
>
>
> > 2. Should zero'ing fudge_factor work? If yes, where should I look to
> remedy
> > the above?
> >
> > I looked upstream for changes that might address this issue but didn't
> see
> > anything.
> >
> > I suspect I can invert my logic and default fudge_factor to some value
> and
> > then override as needed (e.g. 0 for my sparrow platform & 4M for sel4test
> > builds).
>
> This seems fine to me.
>
> >
> > -Sam
> > _______________________________________________
> > Devel mailing list -- [email protected]
> > To unsubscribe send an email to [email protected]
>
_______________________________________________
Devel mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to