Re: [PATCH] KVM: guest_memfd: Disable VMA merging with VM_DONTEXPAND

[David Hildenbrand (Arm)]

On 2/8/26 18:34, Ackerley Tng wrote:
> Ackerley Tng <ackerleytng@xxxxxxxxxx> writes:
>
> > [...snip...]
> >
> > > !thp_vma_allowable_order() must take care of that somehow down in
> > > __thp_vma_allowable_orders(), by checking the file).
> > >
> > > Likely the file_thp_enabled() check is the culprit with
> > > CONFIG_READ_ONLY_THP_FOR_FS?
> > >
> > > Maybe we need a flag to say "even not CONFIG_READ_ONLY_THP_FOR_FS".
> > >
> > > I wonder how we handle that for secretmem. Too late for me, going to bed :)
> >
> > Let me look deeper into this. Thanks!
>
> I trimmed the repro to this:
>
> static void test_guest_memfd_repro(void)
> {
> 	struct kvm_vcpu *vcpu;
> 	uint8_t *unaligned_mem;
> 	struct kvm_vm *vm;
> 	uint8_t *mem;
> 	int fd;
>
> 	vm = __vm_create_shape_with_one_vcpu(VM_SHAPE_DEFAULT, &vcpu, 1, guest_code);
>
> 	fd = vm_create_guest_memfd(vm, SZ_2M * 2, GUEST_MEMFD_FLAG_MMAP |
> GUEST_MEMFD_FLAG_INIT_SHARED);
>
> 	unaligned_mem = mmap(NULL, SZ_2M + SZ_2M, PROT_READ | PROT_WRITE,
> MAP_FIXED | MAP_SHARED, fd, 0);
> 	mem = align_ptr_up(unaligned_mem, SZ_2M);
> 	TEST_ASSERT(((unsigned long)mem & (SZ_2M - 1)) == 0, "returned
> address must be aligned to SZ_2M");
>
> 	TEST_ASSERT_EQ(madvise(mem, SZ_2M, MADV_HUGEPAGE), 0);
>
> 	for (int i = 0; i < SZ_2M; i += SZ_4K)
> 		READ_ONCE(mem[i]);
>
> 	TEST_ASSERT_EQ(madvise(mem, SZ_2M, MADV_COLLAPSE), 0);
>
> 	TEST_ASSERT_EQ(madvise(mem, SZ_2M, MADV_DONTNEED), 0);
>
> 	/* This triggers the WARNing. */
> 	READ_ONCE(mem[0]);
>
> 	munmap(unaligned_mem, SZ_2M * 2);
>
> 	close(fd);
> 	kvm_vm_free(vm);
> }
>
> And tried to replace the fd creation the secretmem equivalent
>
> 	fd = syscall(__NR_memfd_secret, 0);
> 	TEST_ASSERT(fd >= 0, "Couldn't create secretmem fd.");
> 	TEST_ASSERT_EQ(ftruncate(fd, SZ_2M * 2), 0);
>
> Should a guest_memfd selftest be added to cover this?
>
> MADV_COLLAPSE fails with EINVAL, but it does go through to
> hpage_collapse_scan_file() -> collapse_file(), before failing because
> when collapsing the page, copy_mc_highpage() returns > 0.

Just what I suspected. :)

Thanks for digging into the details!

--
Cheers,

David