Re: Questions about mmap

["Chris Friesen" <cfriesen@nortel.com>]

linux-os (Dick Johnson) wrote:
> On Tue, 7 Oct 2008, Thiago Lacerda wrote:
> 
> 
>>Thank you Stefan.... by I'd something more concrete.
>>
>>I'm trying to do like this:
>>
>>//Code of the char device
>>unsigned int **test;
>>static int device_mmap(struct file *filp, struct vm_area_struct *vma) {
>>     printk(KERN_INFO"Calling mmap\n");
>>     vma->vm_flags |= VM_LOCKED;
>>     if(remap_pfn_range(vma, vma->vm_start,
>>virt_to_phys((void*)((unsigned long)test)) >> PAGE_SHIFT, vma->vm_end
>>- vma->vm_start, PAGE_SHARED))
>>               return -EAGAIN;
>>     printk(KERN_INFO"mmap returned\n");
>>     return 0;
>>}
>>
>>static int __init testeInit(void) {
>>/* after creating char device and registering it*/
>>...
>>....
>>test = (unsigned int**) kmalloc(sizeof(unsigned int*)*1024, GFP_KERNEL);
>>int i;
>>for(i = 0; i < 1024; i++)
>>    test[i] = NULL;
>>
>>unsigned int* temp1 = (unsigned int*) kmalloc(sizeof(unsigned int), GFP_KERNEL);
>>(*temp1) = 9;
>>test[0] = temp1;
>>return 0;
>>}
>>
>>and at user program:
>>
>>int main() {
>>   int fd;
>>
>>   unsigned int **mptr;
>>   size_t size = 1024*sizeof(unsigned int*);
>>   fd = open("/dev/myDev", O_RDWR);
>>   if( fd == -1) {
>>       printf("open error...\n");
>>       exit(0);
>>   }
>>
>>   mptr = mmap(0, sizeof(unsigned int*)*1024, PROT_READ | PROT_WRITE,
>>MAP_FILE | MAP_SHARED, fd, 0);
>>   if(mptr == MAP_FAILED) {
>>       printf("mmap() failed\n");
>>       exit(1);
>>   }
>>   printf("teste[0]: %u\n", *mptr[0]);
>>   munmap(mptr, size);
>>   close(fd);
>>   return 0;
>>}
>>
>>It's not working, could you tell me what am I doing wrong? It is for
>>my undergraduate thesis ans it is really driving me mad.
>>
>>I hope that you can help me.
>>
>>best regards.
>>
>>--
>>Thiago de Barros Lacerda
> 
> 
> It looks like you are trying to memory-map a pointer that
> you expect to de-reference in user space. Perhaps you can
> tell us what it is that you are trying to do. You need
> to use copy/to/from_user to copy things from or to the kernel.
> 
> Attempts to dereference kernel data will fail because the
> segments for kernel data or code are not the same as user
> data or code.

I do not think this is entirely true.  I haven't tried it recently, but 
I have reserved a page of kernel memory in a driver and mapped it into 
userspace such that it can be accessed by both.

This can be useful for cases where you want to make continuously-updated 
information available to userspace without the overhead of a syscall.

There can be some gotchas though...last time I tried this on ARM I had 
to explicitly flush the cache on the memory area after writing it in the 
kernel to ensure that userspace saw the updated data since it was 
accessing it via a different virtual address.

Chris