Re: [External] Re: [PATCH 4/8] mshare: selftests: Add test case shared memory

[Anthony Yznaga <anthony.yznaga@oracle.com>]

On 9/1/25 5:50 AM, Yongting Lin wrote:
> On Fri, Aug 29, 2025 at 9:00 AM Anthony Yznaga
> <anthony.yznaga@oracle.com> wrote:
>>
>> Hi Yongting,
>>
>> Thank you for doing this. This is a great start for testing mshare.
>> I do have some comments below.
>>
>> On 8/25/25 7:57 AM, Yongting Lin wrote:
>>> This test case aims to verify the basic functionalities of mshare.
>>>
>>> Create a mshare file and use ioctl to create mapping for host mm
>>> with supportive flags, then create two processes to map mshare file
>>> to their memory spaces, and eventually verify the correctiness
>>> of sharing memory.
>>>
>>> Signed-off-by: Yongting Lin <linyongting@bytedance.com>
>>> ---
>>>    tools/testing/selftests/mshare/basic.c | 81 +++++++++++++++++++++++++-
>>>    1 file changed, 79 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/tools/testing/selftests/mshare/basic.c b/tools/testing/selftests/mshare/basic.c
>>> index 35739b1133f7..2347d30adfee 100644
>>> --- a/tools/testing/selftests/mshare/basic.c
>>> +++ b/tools/testing/selftests/mshare/basic.c
>>> @@ -3,9 +3,86 @@
>>>    #include "../kselftest_harness.h"
>>>    #include "util.c"
>>>
>>> -TEST(basic)
>>> +#define STRING "I am Msharefs"
>>> +
>>> +FIXTURE(basic)
>>> +{
>>> +     char filename[128];
>>> +     size_t align_size;
>>> +     size_t allocate_size;
>>> +};
>>> +
>>> +FIXTURE_VARIANT(basic) {
>>> +     /* decide the time of real mapping size besed on align_size */
>>> +     size_t map_size_time;
>>> +     /* flags for ioctl */
>>> +     int map_flags;
>>> +};
>>> +
>>> +FIXTURE_VARIANT_ADD(basic, ANON_512G) {
>>> +     .map_size_time = 1,
>>> +     .map_flags = MAP_ANONYMOUS | MAP_SHARED | MAP_FIXED,
>>> +};
>>> +
>>> +FIXTURE_VARIANT_ADD(basic, HUGETLB_512G) {
>>> +     .map_size_time = 1,
>>> +     .map_flags = MAP_ANONYMOUS | MAP_HUGETLB | MAP_SHARED | MAP_FIXED,
>>> +};
>>> +
>>> +FIXTURE_VARIANT_ADD(basic, ANON_1T) {
>>> +     .map_size_time = 2,
>>> +     .map_flags = MAP_ANONYMOUS | MAP_SHARED | MAP_FIXED,
>>> +};
>>> +
>>> +FIXTURE_VARIANT_ADD(basic, HUGETLB_1T) {
>>> +     .map_size_time = 2,
>>> +     .map_flags = MAP_ANONYMOUS | MAP_HUGETLB | MAP_SHARED | MAP_FIXED,
>>> +};
>>> +
>>> +FIXTURE_SETUP(basic)
>>>    {
>>> -     printf("Hello mshare\n");
>>> +     int fd;
>>> +
>>> +     self->align_size = mshare_get_info();
>>> +     self->allocate_size = self->align_size * variant->map_size_time;
>>> +
>>> +     fd = create_mshare_file(self->filename, sizeof(self->filename));
>>> +     ftruncate(fd, self->allocate_size);
>>> +
>>> +     ASSERT_EQ(mshare_ioctl_mapping(fd, self->allocate_size, variant->map_flags), 0);
>>
>> The tests should differentiate between how much VA space is allocated to
>> an mshare region (i.e with ftruncate()) and how much memory is allocated
>> within an mshare region through the ioctl. While the bounds of an mshare
>> region need to be aligned to 512 GB, the memory allocated within it does
>> not. Right now the tests will try to map 512 GB or 1 TB of anon or
>> hugetlb memory in an mshare region which will fail on smaller systems to
>> due to insufficient memory. Better to allocate smaller amounts so the
>> tests can run on more systems.
>>
>> Anthony
> 
> I Changed my code to allocate a smaller chunk of memory (i.e.
> 4K/8K/2M/4M), and these tests are passed.
> 
> But I found something different:
> step1:  ftruncate a mshare file to 512G
> step2:  ioctl map 8K
> step3: but after that, I am going to mmap 8K to a process but it
> fails, then I have up to mmap 512G memory to process.

This is correct. For page table sharing to work correctly the entire 
range of the mshare region must be mmap'd into the process even though 
the amount of memory actually mapped within the region via ioctl() may 
be much smaller.

> step4: Accessing the memory within the 8K boundary is fine but get
> segfault after exceling the boundary (as the vma of host mm only holds
> a memory region of 8K)
> 
> Should the mmap region keep consistent with the ioctl map region in
> size? (currently, ioctl map region is 8K, but mmap region is 512G)

It might help to think of the 512G mmap region as a window that gives 
the process visbility or access to the actual memory that is mapped 
within the mshare region.

Anthony


> 
> Yongting
>>
>>> +     close(fd);
>>> +}
>>> +
>>> +FIXTURE_TEARDOWN(basic)
>>> +{
>>> +     ASSERT_EQ(unlink(self->filename), 0);
>>> +}
>>> +
>>> +TEST_F(basic, shared_mem)
>>> +{
>>> +     int fd;
>>> +     void *addr;
>>> +     pid_t pid = fork();
>>> +
>>> +     ASSERT_NE(pid, -1);
>>> +
>>> +     fd = open(self->filename, O_RDWR, 0600);
>>> +     ASSERT_NE(fd, -1);
>>> +
>>> +     addr = mmap(NULL, self->allocate_size, PROT_READ | PROT_WRITE,
>>> +                    MAP_SHARED, fd, 0);
>>> +     ASSERT_NE(addr, MAP_FAILED);
>>> +
>>> +     if (pid == 0) {
>>> +             /* Child process write date the shared memory */
>>> +             memcpy(addr, STRING, sizeof(STRING));
>>> +             exit(0);
>>> +     }
>>> +
>>> +     ASSERT_NE(waitpid(pid, NULL, 0), -1);
>>> +
>>> +     /* Parent process should retrieve the data from the shared memory */
>>> +     ASSERT_EQ(memcmp(addr, STRING, sizeof(STRING)), 0);
>>>    }
>>>
>>>    TEST_HARNESS_MAIN
>>
>