Re: [PATCH v2 3/3] mm/vmalloc: skip vmalloc_dump_obj for non-vmalloc addresses

[Ye Liu <ye.liu@linux.dev>]

在 2026/9/22 20:13, Uladzislau Rezki 写道:
> On Mon, Sep 21, 2026 at 09:17:23PM +0800, Ye Liu wrote:
>> From: Ye Liu <liuye@kylinos.cn>
>>
>> vmalloc_dump_obj() unconditionally searches all vmap nodes even when
>> called with a non-vmalloc address (e.g. a slab or stack pointer from
>> mem_dump_obj()).  Add an is_vmalloc_or_module_addr() check at the
>> entry to avoid the unnecessary per-node trylock and rb-tree traversal.
>>
>> Use is_vmalloc_or_module_addr() rather than is_vmalloc_addr() because
>> module, BPF, and execmem allocations reside in MODULES_VADDR..MODULES_END
>> on x86_64, arm64, and riscv -- outside VMALLOC_START..VMALLOC_END -- but
>> are still tracked in the same vmap_nodes rb-tree.
>>
>> Signed-off-by: Ye Liu <liuye@kylinos.cn>
>> ---
>>  mm/vmalloc.c | 3 +++
>>  1 file changed, 3 insertions(+)
>>
>> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
>> index 30c610f678dc..d8095b558365 100644
>> --- a/mm/vmalloc.c
>> +++ b/mm/vmalloc.c
>> @@ -5277,6 +5277,9 @@ bool vmalloc_dump_obj(void *object)
>>  	unsigned long addr;
>>  	unsigned long nr_pages;
>>  
>> +	if (!is_vmalloc_or_module_addr(object))
>> +		return false;
>> +
>>  	addr = PAGE_ALIGN_DOWN((unsigned long) object);
>>  
>>  	/*
>>
>> -- 
>> 2.25.1
>>
> Do we need this check? If it is not the vmalloc address, we just return
> noting. Another question is why do you want is_vmalloc_or_module_addr()?
> 
> vmalloc_dump_obj() is about VMALLOC_START..VMALLOC_END, IMO.
> 
> There are only two users of it and both rely on the VMALLOC_START..VMALLOC_END
> range:
> 
> <snip>
> *** mm/kasan/report.c:
> print_address_description[403] if (!vmalloc_dump_obj(addr))
> 
> *** mm/util.c:
> mem_dump_obj[1096]             if (vmalloc_dump_obj(object))
> <snip>
> 
> 	if (is_vmalloc_addr(addr)) {
> 		pr_err("The buggy address belongs to a");
> 		if (!vmalloc_dump_obj(addr))
> 			pr_cont(" vmalloc virtual mapping\n");
> 		page = vmalloc_to_page(addr);
> 	}
> 
> and
> 
> <snip>
> 	if (vmalloc_dump_obj(object))
> 		return;
> 
> 	if (is_vmalloc_addr(object))
> 		type = "vmalloc memory";
> <snip>
> 
> --
> Uladzislau Rezki
Hi Uladzislau,                                                        
                                                                      
You're right that the check is not necessary — without it,            
vmalloc_dump_obj() just returns false for non-vmalloc addresses after 
searching the rb-tree, which is fine for a debug path.                
                                                                      
Regarding is_vmalloc_or_module_addr(): the concern was that module/BPF
allocations are also tracked in vmap_nodes rb-tree (they go through   
__vmalloc_node_range with MODULES_VADDR..MODULES_END), so             
is_vmalloc_addr() would filter them out.  But I agree that            
vmalloc_dump_obj() is about VMALLOC_START..VMALLOC_END, and both      
callers already use is_vmalloc_addr() in their logic.                 
                                                                      
I will drop this patch from v3.  The series will be two patches: the  
alignment fix and the cross-zone lookup fix with the shared           
find_vmap_area_lock() helper.                                         

-- 
Thanks,
Ye Liu