Re: [PATCH] KVM: selftests: Speed up set_memory_region_test

[Zenghui Yu]

On 2021/4/26 21:01, Vitaly Kuznetsov wrote:
> After commit 4fc096a99e01 ("KVM: Raise the maximum number of user memslots")
> set_memory_region_test may take too long, reports are that the default
> timeout value we have (120s) may not be enough even on a physical host.
>
> Speed things up a bit by throwing away vm_userspace_mem_region_add() usage
> from test_add_max_memory_regions(), we don't really need to do the majority
> of the stuff it does for the sake of this test.
>
> On my AMD EPYC 7401P, # time ./set_memory_region_test
> pre-patch:
>  Testing KVM_RUN with zero added memory regions
>  Allowed number of memory slots: 32764
>  Adding slots 0..32763, each memory region with 2048K size
>  Testing MOVE of in-use region, 10 loops
>  Testing DELETE of in-use region, 10 loops
>
>  real	0m44.917s
>  user	0m7.416s
>  sys	0m34.601s
>
> post-patch:
>  Testing KVM_RUN with zero added memory regions
>  Allowed number of memory slots: 32764
>  Adding slots 0..32763, each memory region with 2048K size
>  Testing MOVE of in-use region, 10 loops
>  Testing DELETE of in-use region, 10 loops
>
>  real	0m20.714s
>  user	0m0.109s
>  sys	0m18.359s
>
> Reported-by: kernel test robot <oliver.sang@xxxxxxxxx>
> Signed-off-by: Vitaly Kuznetsov <vkuznets@xxxxxxxxxx>

I've seen the failure on my arm64 server, # ./set_memory_region_test

Allowed number of memory slots: 32767
Adding slots 0..32766, each memory region with 2048K size
==== Test Assertion Failure ====
  set_memory_region_test.c:391: ret == 0
  pid=42696 tid=42696 errno=22 - Invalid argument
     1	0x00000000004015a7: test_add_max_memory_regions at 
set_memory_region_test.c:389
     2	 (inlined by) main at set_memory_region_test.c:426
     3	0x0000ffffb7c63bdf: ?? ??:0
     4	0x00000000004016db: _start at :?
  KVM_SET_USER_MEMORY_REGION IOCTL failed,
  rc: -1 errno: 22 slot: 2624

> +	mem = mmap(NULL, MEM_REGION_SIZE * max_mem_slots + alignment,

The problem is that max_mem_slots is declared as uint32_t, the result
of (MEM_REGION_SIZE * max_mem_slots) is unexpectedly truncated to be
0xffe00000.

> +		   PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
> +	TEST_ASSERT(mem != MAP_FAILED, "Failed to mmap() host");
> +	mem_aligned = (void *)(((size_t) mem + alignment - 1) & ~(alignment - 1));
> +
>  	for (slot = 0; slot < max_mem_slots; slot++) {
> -		vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS,
> -					    guest_addr, slot, mem_reg_npages,
> -					    0);
> -		guest_addr += MEM_REGION_SIZE;
> +		ret = test_memory_region_add(vm, mem_aligned +
> +					     ((uint64_t)slot * MEM_REGION_SIZE),

These unmapped VAs got caught by access_ok() checker in
__kvm_set_memory_region() as they happen to go beyond the task's
address space on arm64. Casting max_mem_slots to size_t in both
mmap() and munmap() fixes the issue for me.

Thanks,
Zenghui