On 10/25/24 at 05:11pm, David Hildenbrand wrote:

This is based on "[PATCH v3 0/7] virtio-mem: s390 support" [1], which adds
virtio-mem support on s390.

The only "different than everything else" thing about virtio-mem on s390
is kdump: The crash (2nd) kernel allocates+prepares the elfcore hdr
during fs_init()->vmcore_init()->elfcorehdr_alloc(). Consequently, the
crash kernel must detect memory ranges of the crashed/panicked kernel to
include via PT_LOAD in the vmcore.

On other architectures, all RAM regions (boot + hotplugged) can easily be
observed on the old (to crash) kernel (e.g., using /proc/iomem) to create
the elfcore hdr.

On s390, information about "ordinary" memory (heh, "storage") can be
obtained by querying the hypervisor/ultravisor via SCLP/diag260, and
that information is stored early during boot in the "physmem" memblock
data structure.

But virtio-mem memory is always detected by as device driver, which is
usually build as a module. So in the crash kernel, this memory can only be
properly detected once the virtio-mem driver started up.

The virtio-mem driver already supports the "kdump mode", where it won't
hotplug any memory but instead queries the device to implement the
pfn_is_ram() callback, to avoid reading unplugged memory holes when reading
the vmcore.

With this series, if the virtio-mem driver is included in the kdump
initrd -- which dracut already takes care of under Fedora/RHEL -- it will
now detect the device RAM ranges on s390 once it probes the devices, to add
them to the vmcore using the same callback mechanism we already have for
pfn_is_ram().

To add these device RAM ranges to the vmcore ("patch the vmcore"), we will
add new PT_LOAD entries that describe these memory ranges, and update
all offsets vmcore size so it is all consistent.

Note that makedumfile is shaky with v6.12-rcX, I made the "obvious" things
(e.g., free page detection) work again while testing as documented in [2].

Creating the dumps using makedumpfile seems to work fine, and the
dump regions (PT_LOAD) are as expected. I yet have to check in more detail
if the created dumps are good (IOW, the right memory was dumped, but it
looks like makedumpfile reads the right memory when interpreting the
kernel data structures, which is promising).

Patch #1 -- #6 are vmcore preparations and cleanups

Thanks for CC-ing me, I will review the patch 1-6, vmcore part next
week.