On Mon, Jan 26, 2026 at 02:28:30PM +0100, Pratyush Yadav wrote:
> > On Wed, Jan 21, 2026 at 06:50:38AM -0800, Breno Leitao wrote:
> >> +static __init int kho_populate_kexec_metadata(void)
> >> +{
> >> + struct kho_kexec_metadata *metadata;
> >> + int err;
> >> +
> >> + metadata = kho_alloc_preserve(sizeof(*metadata));
> >> + if (IS_ERR(metadata))
> >> + return PTR_ERR(metadata);
> >> +
> >> + strscpy(metadata->previous_release, init_uts_ns.name.release,
> >> + sizeof(metadata->previous_release));
> >> + /* kho_in.kexec_count is set to 0 on cold boot */
> >> + metadata->kexec_count = kho_in.kexec_count + 1;
> >> +
> >> + err = kho_add_subtree(KHO_METADATA_NODE_NAME, metadata);
> >
> > There is a hidden bug in here when CONFIG_KEXEC_HANDOVER_DEBUGFS is set.
>
> Good catch!
>
> >
> > kho_add_subtree() expects a fdt as the second argument, and we are
> > passing a pure C struct. That works fine, except for debugfs, which
> > does:
> >
> > 1. kho_add_subtree() calls kho_debugfs_fdt_add()
> > 2. kho_debugfs_fdt_add() calls __kho_debugfs_fdt_add()
> > 3. __kho_debugfs_fdt_add() executes fdt_totalsize(fdt)
> >
> > The fdt_totalsize() macro reads bytes 4-7 of the input as a big-endian u32,
> > and
> > this will hit struct kho_kexec_metadata, given I am passing a C struct
> > instead
> > of a FDT.
> >
> > struct kho_kexec_metadata {
> > char previous_release[__NEW_UTS_LEN + 1]; // 65 bytes
> > u32 kexec_count;
> > } __packed;
> >
> > Bytes 4-7 would be characters from previous_release (e.g., "0-rc" from
> > "6.19.0-rc4..."). Interpreted as big-endian u32, this gives a garbage size
> > value.
> >
> > The alternatives I see here are:
> >
> > 1) Come back to FDT instead of plain C struct, similarly to the previous
> > version [1]
> > 2) Created some helpers to treat C struct fields specially just for this
> > feature, and we can do it later if we have more users.
> > 3) Move this kexec_metadata to work on top of LUO (similarly to memfd), but
> > that would be an unnecessary dependency just to have this
> > kexec_metadata.
> >
> > That said, for the next version, I am coming back to to FDT.
>
> Please, no. Don't go back to it just for the sake of this bug.
>
> I think KHO's assumption that the subtree will always point to an FDT is
> broken, and we should fix that. I think KHO should expose the blob of
> serialized data and let userspace figure out what the format is and how
> to decode it.
>
> To do that, we would need to update kho_add_subtree() to take a size
> parameter from callers, and pass that down to debugfs code. I count 3
> callers of kho_add_subtree() - memblock, LUO, and test_kho. I think all
> 3 should be fairly easy to update, but I am happy to help out if you
> need.
Sure, let me hack and see what I can get here.
Thanks for the direction,
--breno
