Re: [PATCH] ppc/fadump: collect dump if the collected size is lesser than reserved

From: Sourabh Jain

Date: Wed Jul 29 2026 - 02:33:34 EST

On 29/07/26 11:36, Shivang Upadhyay wrote:

On Wed, 2026-07-29 at 09:14 +0530, Sourabh Jain wrote:

On 22/07/26 14:41, Shivang Upadhyay wrote:

On Sun, 2026-07-19 at 12:39 +0530, Sourabh Jain wrote:

Could you please reword the commit title to make it a bit
clearer?

On 14/07/26 23:00, Shivang Upadhyay wrote:

When a machine is subjected to CPUs add/remove, using dlpar
operations, the number of collected CPU_NOTES can change.
As per PAPR, collected dump size should not be more than
allocated size. Reflecting the same in source.

Can you add more details about the problem you are trying to
solve
with this patch and how.

Adding the error message and scenario would be really helpful in
understanding the problem.

Hi Sourabh,

When a qemu ppc machine is booted with fadump=on and `-
smp=x,maxcpus=y`, on the panic kernel, /proc/vmcore is not
generated
because dump_bytes and source_len does'nt match for CPU_STATE_DATA
region in fadump, then we just give up on parsing rest of the data.

Can you add Closes tag if it is reported upstream and if possible
fixes
tag too.

Signed-off-by: Shivang Upadhyay <shivangu@xxxxxxxxxxxxx>
---
   arch/powerpc/platforms/pseries/rtas-fadump.c | 3 ++-
   1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/arch/powerpc/platforms/pseries/rtas-fadump.c
b/arch/powerpc/platforms/pseries/rtas-fadump.c
index 3bb4ac2ab6cc..19a5adaf326b 100644
--- a/arch/powerpc/platforms/pseries/rtas-fadump.c
+++ b/arch/powerpc/platforms/pseries/rtas-fadump.c
@@ -469,7 +469,8 @@ static int __init
rtas_fadump_process(struct
fw_dump *fadump_conf)
    pr_err("Dump taken by platform
is
not valid (%d)\n", i);
    rc = -EINVAL;
    }
- if (fdm_active->rgn[i].bytes_dumped !=
fdm_active->rgn[i].source_len) {
+ if (be64_to_cpu(fdm_active-

rgn[i].bytes_dumped)

+     > be64_to_cpu(fdm_active-

rgn[i].source_len)) {

Can you please share your observations about `bytes_dump` for
both
QEMU
and a
real system (LPAR) where the number of online CPUs is not equal
to
the
maximum

I have the following observation.

I booted LPAR with 8 cpus. After crashing it I saw that fadump
CPU_STATE_DATA has notes for total 16 cpus, and only top 8 notes
have
valid entries.when trying with 16, I see all NOTES have entried
filled.

Yes even I noticed the same on a LPAR with (Min=1 Desired=1 Max=2
with
SMT=8 CPUs) maxcpus as 16 and online CPUs as 8.

[ 0.036128] rtas fadump: --------CPU State Data------------
[ 0.036129] rtas fadump: Magic Number: 5245475341564500
[ 0.036131] rtas fadump: NumCpuOffset: 1c
[ 0.036132] rtas fadump: NumCpus : 16
[ 0.036135] fadump: Allocated buffer for cpu notes of size 65536
at
0xc000000006a30000
[ 0.036138] rtas fadump: Reading register data for cpu 0...
[ 0.036173] rtas fadump: Reading register data for cpu 1...
[ 0.036178] rtas fadump: Reading register data for cpu 2...
[ 0.036209] rtas fadump: Reading register data for cpu 3...
[ 0.036240] rtas fadump: Reading register data for cpu 4...
[ 0.036272] rtas fadump: Reading register data for cpu 5...
[ 0.036304] rtas fadump: Reading register data for cpu 6...
[ 0.036335] rtas fadump: Reading register data for cpu 7...
[ 0.036390] rtas fadump: Updating elfcore header
(c000000006a20000)
with cpu notes

The NumCpus is populated to be 16 CPUs by the firmware (RTAS/PHYP)
even
though only 8 CPUs were online.

Kernel avoid processing reg entries of CPUs which were offline using
below condition.
code snippet from rtas_fadump_build_cpu_notes()/rtas-faudmp.c

   if (fdh && !cpumask_test_cpu(cpu, &fdh->cpu_mask)) {
   RTAS_FADUMP_SKIP_TO_NEXT_CPU(reg_entry);
   continue;
   }

But when I removed the avoid condition kernel failed to process the
reg
entries and
below logs were printed.

[ 0.037877] rtas fadump: --------CPU State Data------------
[ 0.037879] rtas fadump: Magic Number: 5245475341564500
[ 0.037881] rtas fadump: NumCpuOffset: 1c
[ 0.037883] rtas fadump: NumCpus : 16
[ 0.037889] fadump: Allocated buffer for cpu notes of size 65536
at
0xc000000007a50000
[ 0.037891] rtas fadump: Reading register data for cpu 0...
[ 0.037937] rtas fadump: Reading register data for cpu 1...
[ 0.037974] rtas fadump: Reading register data for cpu 2...
[ 0.038009] rtas fadump: Reading register data for cpu 3...
[ 0.038044] rtas fadump: Reading register data for cpu 4...
[ 0.038079] rtas fadump: Reading register data for cpu 5...
[ 0.038082] rtas fadump: Reading register data for cpu 6...
[ 0.038117] rtas fadump: Reading register data for cpu 7...
[ 0.038155] rtas fadump: CPU 8 was offline
[ 0.038157] rtas fadump: Reading register data for cpu 8...
[ 0.038195] rtas fadump: CPU 10 was offline
[ 0.038197] rtas fadump: Reading register data for cpu 10...
[ 0.038235] rtas fadump: CPU 12 was offline
[ 0.038236] rtas fadump: Reading register data for cpu 12...
[ 0.038273] rtas fadump: CPU 14 was offline
[ 0.038275] rtas fadump: Reading register data for cpu 14...
[ 0.038311] rtas fadump: Unable to read CPU state data
[ 0.038319] fadump: Invalidating firmware-assisted dump
registration
[ 0.038370] rtas fadump: Firmware busy during fadump invalidate,
waiting 1ms (total 0ms)
[ 0.039403] fadump: reserved_memory_range[0]
[0x00000040000000-0x000000800305c7], 0x400305c8 bytes
[ 0.039408] fadump: freeing reserved memory (0x80030000 -
0xa00000000)

For some reason, only CPUs with even-numbered IDs were present from
offline CPUs list. This suggests that although
NumCpus was 16, only 8 CPUs had valid reg entries. I think we should
discuss this case with the firmware team before
finalizing the solution.

Thats weird, I remember seeing entries for all of the maxcpus. Maybe I
put my prints differently, Can you share your debug patch source?

diff --git a/arch/powerpc/kernel/fadump.c b/arch/powerpc/kernel/fadump.c
index 7f79c9aea..ee1ad6493 100644
--- a/arch/powerpc/kernel/fadump.c
+++ b/arch/powerpc/kernel/fadump.c
@@ -11,7 +11,7 @@
* Author: Mahesh Salgaonkar <mahesh@xxxxxxxxxxxxxxxxxx>
*/

-#undef DEBUG
+#define DEBUG
#define pr_fmt(fmt) "fadump: " fmt

#include <linux/string.h>
diff --git a/arch/powerpc/platforms/pseries/rtas-fadump.c b/arch/powerpc/platforms/pseries/rtas-fadump.c
index 3bb4ac2ab..55271fbbf 100644
--- a/arch/powerpc/platforms/pseries/rtas-fadump.c
+++ b/arch/powerpc/platforms/pseries/rtas-fadump.c
@@ -6,6 +6,8 @@
* Copyright 2019, Hari Bathini, IBM Corporation.
*/

+#define DEBUG
+
#define pr_fmt(fmt) "rtas fadump: " fmt

#include <linux/string.h>
@@ -422,7 +424,7 @@ static int __init rtas_fadump_build_cpu_notes(struct fw_dump *fadump_conf)
RTAS_FADUMP_CPU_ID_MASK);
if (fdh && !cpumask_test_cpu(cpu, &fdh->cpu_mask)) {
RTAS_FADUMP_SKIP_TO_NEXT_CPU(reg_entry);
- continue;
+ pr_debug("CPU %d was offline\n", cpu);
}
pr_debug("Reading register data for cpu %d...\n", cpu);
if (fdh && fdh->crashing_cpu == cpu) {

On the other hand when I increased the CPUs count on the same system
(Min=2 Desired=4 and max=8 with SMT 8) the
system was booted with 32 possible CPUs instead of 64 CPUs. And
NumCpus
was 32 in fadump kernel.

I previously proposed this fake entries fix in qemu [1]. But that
is
not compliant with PAPR, which states that cpu notes should be
collected for current_cpus only.

Yes, that solution does not appear to be PAPR-compliant. However,
based
on our experiments, it is unclear why the firmware reports NumCpus as
16
when only 8 CPUs are actually online.

I think fw's reason for doing for maxcpus would be same as our case
also, CPU hotplug. It would be hard to handle variable cpus in dump, so
why not just keep entries for all?

I think there is some inconsistency here. The firmware reports 16 CPUs in
NumCpus, but the kernel fails when it tries to read the reg entries for the offline CPUs.

I'm not sure where the problem is yet. Could you dump the Register Save Area for
a case where maxcpus is greater than the number of online CPUs? Then we can
check what data is present for the offline CPUs.

As per current states of things, reservation is always done for
maxcpus, (LPAR and QEMU both), NOTES collections is done for
current
cpus on QEMU, and maxcpus on LPARs.

My proposal here is that we make the check for bytes_dumped and
source_len, less restrictive to let QEMU also generate
/proc/vmcore,
when current_cpus != maxcpus.

By the way, the changes proposed in this patch also apply to other
regions,
such as HPTE and REAL_MODE. That is not the intended behavior, right?

Yes, I made it this way by reading the dump_bytes and source_len
relation. PAPR only says that dumped_bytes should not exceed the
source_len. My change reflects the same restriction in sw. Also along
with this lengths check, we have firmware reported error flag also,
which should take care of overflow (Qemu handles it like this [1]).

But the firmware not copying the entire Real Mode memory is also a problem for
the kernel, isn't it?

- Sourabh Jain

[1]
https://elixir.bootlin.com/qemu/v11.1.0-rc1/source/hw/ppc/spapr_fadump.c#L263

~Shivang.