On 09/04/24 at 02:55pm, Sourabh Jain wrote:

Hello Baoquan,

On 30/08/24 16:47, Baoquan He wrote:

quoted

On 08/20/24 at 12:10pm, Sourabh Jain wrote:

quoted

Hello Baoquan,

......snip... 

quoted

quoted

2. A patch to return early from the `crash_handle_hotplug_event()` function
if `kexec_in_progress` is
    set to True. This is essentially my original patch.

There's a race gap between the kexec_in_progress checking and the
setting it to true which Michael has mentioned.

The window where kernel is holding kexec_lock to do kexec boot
but kexec_in_progress is yet not set to True.

If kernel needs to handle crash hotplug event, the function
crash_handle_hotplug_event()  will not get the kexec_lock and
error out by printing error message about not able to update
kdump image.

But you wanted to avoid the erroring out if it's being in
kernel_kexec().  Now you are seeing at least one the noising 
message, aren't you?

I think it should be fine. Given that lock is already taken for
kexec kernel boot.

Am I missing something major?

quoted

That's why I think
maybe checking kexec_in_progress after failing to retriving
__kexec_lock is a little better, not very sure.

Try for kexec lock before kexec_in_progress check will not solve
the original problem this patch trying to solve.

You proposed the below changes earlier:

-	if (!kexec_trylock()) {
+	if (!kexec_trylock() && kexec_in_progress) {
 		pr_info("kexec_trylock() failed, elfcorehdr may be inaccurate\n");
 		crash_hotplug_unlock();

Ah, I meant as below, but wrote it mistakenly.

diff --git a/kernel/crash_core.c b/kernel/crash_core.c
index 63cf89393c6e..e7c7aa761f46 100644
--- a/kernel/crash_core.c
+++ b/kernel/crash_core.c

@@ -504,7 +504,7 @@ int crash_check_hotplug_support(void)
 
 	crash_hotplug_lock();
 	/* Obtain lock while reading crash information */
-	if (!kexec_trylock()) {
+	if (!kexec_trylock() && !kexec_in_progress) {
 		pr_info("kexec_trylock() failed, elfcorehdr may be inaccurate\n");
 		crash_hotplug_unlock();
 		return 0;

Once the kexec_in_progress is set to True there is no way one can get
kexec_lock. So kexec_trylock() before kexec_in_progress is not helpful
for the problem I am trying to solve.

With your patch, you could still get the error message if the race gap
exist. With above change, you won't get it. Please correct me if I am
wrong.