Re: [PATCH] mm/sparsemem: fix race in accessing memory_section->usage

[Alexander Potapenko]

Cc: "Paul E. McKenney" <paulmck@xxxxxxxxxx>
Cc: Marco Elver <elver@xxxxxxxxxx>
Cc: Dmitry Vyukov <dvyukov@xxxxxxxxxx>
Cc: kasan-dev@xxxxxxxxxxxxxxxx
Cc: Ilya Leoshkevich <iii@xxxxxxxxxxxxx>
Cc: Nicholas Miehlbradt <nicholas@xxxxxxxxxxxxx>

Hi folks,

(adding KMSAN reviewers and IBM people who are currently porting KMSAN to other
architectures, plus Paul for his opinion on refactoring RCU)

this patch broke x86 KMSAN in a subtle way.

For every memory access in the code instrumented by KMSAN we call
kmsan_get_metadata() to obtain the metadata for the memory being accessed. For
virtual memory the metadata pointers are stored in the corresponding `struct
page`, therefore we need to call virt_to_page() to get them.

According to the comment in arch/x86/include/asm/page.h, virt_to_page(kaddr)
returns a valid pointer iff virt_addr_valid(kaddr) is true, so KMSAN needs to
call virt_addr_valid() as well.

To avoid recursion, kmsan_get_metadata() must not call instrumented code,
therefore ./arch/x86/include/asm/kmsan.h forks parts of arch/x86/mm/physaddr.c
to check whether a virtual address is valid or not.

But the introduction of rcu_read_lock() to pfn_valid() added instrumented RCU
API calls to virt_to_page_or_null(), which is called by kmsan_get_metadata(),
so there is an infinite recursion now. I do not think it is correct to stop that
recursion by doing kmsan_enter_runtime()/kmsan_exit_runtime() in
kmsan_get_metadata(): that would prevent instrumented functions called from
within the runtime from tracking the shadow values, which might introduce false
positives.

I am currently looking into inlining __rcu_read_lock()/__rcu_read_unlock(), into
KMSAN code to prevent it from being instrumented, but that might require factoring
out parts of kernel/rcu/tree_plugin.h into a non-private header. Do you think this
is feasible?

Another option is to cut some edges in the code calling virt_to_page(). First,
my observation is that virt_addr_valid() is quite rare in the kernel code, i.e.
not all cases of calling virt_to_page() are covered with it. Second, every
memory access to KMSAN metadata residing in virt_to_page(kaddr)->shadow always
accompanies an access to `kaddr` itself, so if there is a race on a PFN then
the access to `kaddr` will probably also trigger a fault. Third, KMSAN metadata
accesses are inherently non-atomic, and even if we ensure pfn_valid() is
returning a consistent value for a single memory access, calling it twice may
already return different results.

Considering the above, how bad would it be to drop synchronization for KMSAN's
version of pfn_valid() called from kmsan_virt_addr_valid()?