Currently, the mem_dump_obj() is invoked in call_rcu(), the
call_rcu() is maybe invoked in non-preemptive code segment,
for object allocated from vmalloc(), the following scenarios
may occur:

        CPU 0
tasks context
   spin_lock(&vmap_area_lock)
          Interrupt context
              call_rcu()
                mem_dump_obj
                  vmalloc_dump_obj
                    spin_lock(&vmap_area_lock) <--deadlock

and for PREEMPT-RT kernel, the spinlock will convert to sleepable
lock, so the vmap_area_lock spinlock not allowed to get in non-preemptive
code segment. therefore, this commit make the vmalloc_dump_obj() call in
a preemptible context.

Signed-off-by: Zqiang <qiang1.zhang@intel.com>
---
 v1->v2:
 Add IS_ENABLED(CONFIG_PREEMPT_RT) check.

 mm/util.c    | 4 +++-
 mm/vmalloc.c | 4 ++++
 2 files changed, 7 insertions(+), 1 deletion(-)

diff --git a/mm/util.c b/mm/util.c
index 12984e76767e..2b0222a728cc 100644
--- a/mm/util.c
+++ b/mm/util.c
@@ -1128,7 +1128,9 @@ void mem_dump_obj(void *object)
 		return;
 
 	if (virt_addr_valid(object))
-		type = "non-slab/vmalloc memory";
+		type = "non-slab memory";
+	else if (is_vmalloc_addr(object))
+		type = "vmalloc memory";
 	else if (object == NULL)
 		type = "NULL pointer";
 	else if (object == ZERO_SIZE_PTR)
diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index ccaa461998f3..018e417b12d6 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -4034,6 +4034,10 @@ bool vmalloc_dump_obj(void *object)
 	struct vm_struct *vm;
 	void *objp = (void *)PAGE_ALIGN((unsigned long)object);
 
+	if (!is_vmalloc_addr(objp) || ((IS_ENABLED(CONFIG_PREEMPT_RT) &&
+				!preemptible()) || in_interrupt()))
+		return false;
+
 	vm = find_vm_area(objp);
 	if (!vm)
 		return false;
-- 
2.25.1

On Thu, 17 Nov 2022 19:25:20 +0800 Zqiang <qiang1.zhang@intel.com> wrote:

> Currently, the mem_dump_obj() is invoked in call_rcu(), the
> call_rcu() is maybe invoked in non-preemptive code segment,
> for object allocated from vmalloc(), the following scenarios
> may occur:
> 
>         CPU 0
> tasks context
>    spin_lock(&vmap_area_lock)
>           Interrupt context
>               call_rcu()
>                 mem_dump_obj
>                   vmalloc_dump_obj
>                     spin_lock(&vmap_area_lock) <--deadlock
> 
> and for PREEMPT-RT kernel, the spinlock will convert to sleepable
> lock, so the vmap_area_lock spinlock not allowed to get in non-preemptive
> code segment. therefore, this commit make the vmalloc_dump_obj() call in
> a preemptible context.
> 
> ...
>
> --- a/mm/util.c
> +++ b/mm/util.c
> @@ -1128,7 +1128,9 @@ void mem_dump_obj(void *object)
>  		return;
>  
>  	if (virt_addr_valid(object))
> -		type = "non-slab/vmalloc memory";
> +		type = "non-slab memory";
> +	else if (is_vmalloc_addr(object))
> +		type = "vmalloc memory";
>  	else if (object == NULL)
>  		type = "NULL pointer";
>  	else if (object == ZERO_SIZE_PTR)
> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> index ccaa461998f3..018e417b12d6 100644
> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -4034,6 +4034,10 @@ bool vmalloc_dump_obj(void *object)
>  	struct vm_struct *vm;
>  	void *objp = (void *)PAGE_ALIGN((unsigned long)object);
>  
> +	if (!is_vmalloc_addr(objp) || ((IS_ENABLED(CONFIG_PREEMPT_RT) &&
> +				!preemptible()) || in_interrupt()))
> +		return false;
> +
>  	vm = find_vm_area(objp);
>  	if (!vm)
>  		return false;

I suggest this be restructured so we can comment each test:

	/* comment goes here */
	if (!is_vmalloc_addr(objp))
		return false;

	/* comment goes here */
	if (IS_ENABLED(CONFIG_PREEMPT_RT) && !preemptible())
		return false;

	/* comment goes here */
	if (in_interrupt()))
		return false;

Where each comment carefully explains why we're performing the test. 
It will generate the same code.