userspace-kernel shared memory with mmap works strange on multi-thread - linux

Question

I'm currently making shared memory between userspace to kernel.
The following code is simple version of my mmap method.
for (j = 0; j < 3; j++){
RxFrame[j] = kmalloc(4096*6, GFP_KERNEL);
npages = (4096*6 - 1)/(PAGE_SIZE) + 1;
for(i = 0; i < npages*PAGE_SIZE; i += PAGE_SIZE){
printk("RxFrame[%d] PAGE_SIZE is %d\n", j, PAGE_SIZE);
SetPageReserved(virt_to_page(RxFrame[j] + i));
}
}
static int my_mmap(struct file *filp, struct vm_area_struct *vma){
unsigned long pfn = virt_to_phys(RxFrame[index]);
ret = remap_pfn_range(vma, vma->vm_start, pfn>>PAGE_SHIFT, len, vma->vm_page_prot);
if (ret < 0) {
pr_err("could not map the address area\n");
return -EIO;
}
return 0;
}
And following is my userspace application code which mmap used.
FRAME *RxFrame[3];
for(int i = 0; i < 3; i++){
if( (RxFrame[i] = mmap(0, 4096*6, PROT_WRITE | PROT_READ, MAP_SHARED, fd[i], 0)) == NULL ){
printf("mmap error\n");
return 0;
}
}
auto DoReceive = [&](const int index){
// read RxFrame[index][0].data
}
for(int i = 0; i < 3; i++){
thread[i] = std::thread(DoReceive , i);
usleep(1000);
}
My purpose was making three kernel buffers which dedicated data(different data will store in different kernel buffer) will store and use them with mmap in userspace with multi-thread.
when just one kernel buffer is used(data stored in just one kernel buffer), I can read RxFrame[index][0].data very well.
But when three kernel buffer is used(different data stored in separate kernel buffer), I can only read the first transmitted data.. for example if RxFrame[0]'s buffer received data, I can read RxFrame[0][0].data but can't read RxFrame[1][0].data and RxFrame[2][0].data..
Strange thing is that when received data is quite a lot at few moment, I can read RxFrame[1][0].data and RxFrame[2][0].data which looks like cached result!
I used volatile and other mmap option but I can't figure out why my mapped memory works like cached in multi-thread.
Any help will be thanks!
userspace-kernel shared memory with mmap works strange on multi-thread