Re: [PATCH] don't record leak information on allocations done between kmemleak_init and kmemleak_late_init

[Catalin Marinas]

(as you noticed, don't post html as they usually get rejected from
lists)

On Wed, Jan 03, 2024 at 06:20:16AM -0500, Christoph von Recklinghausen wrote:
> On 1/2/24 15:07, Catalin Marinas wrote:
> > On Tue, Jan 02, 2024 at 10:34:28AM -0500, Chris von Recklinghausen wrote:
> > > If an object is allocated after kmemleak_init is called but before
> > > kmemleak_late_init is called, calls to kmemleak_not_leak or
> > > kmemleak_ignore on the object don't prevent a scan from reporting the
> > > object as a leak.
> > This may be true but what is the reason for this? Can you give some
> > example of false positives you get?
> 
> In centos-stream-9 on s390x I get the following complaint:
> 
> WARN:(libsan.host.linux) Found kernel memory leak:
> unreferenced object 0x1bff7fffb30000 (size 65536):
> comm "swapper/0", pid 0, jiffies 4294937297 (age 76.530s)
> hex dump (first 32 bytes):
> 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
> 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
> backtrace:
> [<00000000eda98345>] __vmalloc_node_range+0x29a/0x360
> [<00000000e3051c75>] __vmalloc_node+0x9e/0xd0
> [<00000000a5dd11b7>] stack_alloc+0x38/0x50
> [<0000000081096e42>] smp_reinit_ipl_cpu+0xf8/0x3f8
> [<00000000ee13aae5>] arch_call_rest_init+0x22/0x100
> [<00000000b37567c9>] start_kernel+0x44c/0x460
> [<00000000548d9080>] startup_continue+0x30/0x50
> 
> 
> In arch/s390/kernel/setup.c we have
> 
> unsigned long stack_alloc(void)
> {
> #ifdef CONFIG_VMAP_STACK
>         return (unsigned long)__vmalloc_node(THREAD_SIZE, THREAD_SIZE,
>                         THREADINFO_GFP, NUMA_NO_NODE,
>                         __builtin_return_address(0));
> #else
>         return __get_free_pages(GFP_KERNEL, THREAD_SIZE_ORDER);
> #endif
> }

I guess that's an older kernel as arch_call_rest_init() is no longer in
mainline.

Mainline stack_alloc() has a kmemleak_not_leak() call here with an
explanation in the commit log (it should have been added as a comment in
the code), 436fc4feeabb ("s390: add kmemleak annotation in
stack_alloc()"):

    kmemleak with enabled auto scanning reports that our stack allocation is
    lost. This is because we're saving the pointer + STACK_INIT_OFFSET to
    lowcore. When kmemleak now scans the objects, it thinks that this one is
    lost because it can't find a corresponding pointer.

Does this commit not fix it for you? It looks like it did the trick in
mainline. Late kmemleak initialisation should not interfere unless you
have a very old kernel and we had some bugs with tracking these (before
we introduced the mem_pool[] array in kmemleak for early allocations).

> void __init arch_call_rest_init(void)
> {
>         unsigned long stack;
> 
>         smp_reinit_ipl_cpu();
>         stack = stack_alloc();
>         if (!stack)
>                 panic("Couldn't allocate kernel stack");
>         current->stack = (void *) stack;
> #ifdef CONFIG_VMAP_STACK
>         current->stack_vm_area = (void *) stack;
> #endif

In mainline at least, stack_vm_area is a struct vm_struct, so it
shouldn't be assigned the actual stack pointer (but maybe that's not the
case in your version, I haven't checked the history).

-- 
Catalin