A question about iommu, passing to the device user data containing pointers

May 2, 2022

      Hello linux experts,

I'm almost sure my understanding is correct but hope someone could confirm
me if it's really correct.
Suppose a user program wants to pass to a device some data structures
connected to each other using pointers and the device processes the data
with dma capability(meaning it accesses the memory on its own). Here is a
simple example. I want to add each member of array a and b and put it in
array c and I pass the address of the struct which contains the addresses of
inputs and output arrays. (This is a code for this question and can have
some bugs but it's enough for my question. )  

    struct _myargs {
    int *a;  // input pointer
    int *b;  // input pointer
    int *c;  // output pointer
    } myargs;

    int main()
    {
    int *a, *b, *c; 
    a = malloc(100*sizeof(int));
    b = malloc(100*sizeof(int));
    c = malloc(100*sizeof(int)); 

    ... fill inputs ...

    myargs.a = a;
    myargs.b = b;
    myargs.c = c;
    ...
    ioctl(fd, AddIt, &myargs);
    ...
    }

 Then, in the driver, I allocate the input and output buffers in kernel
space and I do copy_from_user to copy the input data to kernel space input
buffers. (and later do copy_to_user for the buffer c).   

    my_ioctl(struct file *file, unsigned int cmd, unsigned long args_usr)
    {
    struct _myargs *args_ker;
    dma_addr_t args_dev;
    int *a_ker, *b_ker, *c_ker;
    dma_addr a_dev, b_dev, c_dev;

    args = dma_alloc_coherent(dev, sizeof(struct _myargs), &args_dev,
GFP_KERNEL);

    switch(cmd) {
    case(AddIt):
    copy_from_user(args_ker, args_usr, sizeof(struct _myargs));
    ak = dma_alloc_coherent(dev, 100*sizeof(int), &a_dev, GFP_KERNEL);
    bk = dma_alloc_coherent(dev, 100*sizeof(int), &b_dev, GFP_KERNEL);
    ck = dma_alloc_coherent(dev, 100*sizeof(int), &c_dev, GFP_KERNEL);
    copy_from_user(a_ker, args_ker.a, 100*sizeof(int));
    copy_from_user(b_ker, args_ker.b, 100*sizeof(int));

    args_ker.a = a_dev;
    args_ker.b = b_dev;
    args_ker.c = c_dev;

    writel(dev, &args_dev, REG_ARGS); // set args
    writel(dev, 1, REG_TRIGGER);   // trigger hardware

    ...
    }
    }

So the a, b arrays have been copied to kernel space address a_ker and b_ker
but when I pass the args struct to the device, I used virtual io address
(dma_addr_t) a_dev, b_dev and c_dev.  And I want the device hardware to read
the args struct from memory and use the address there to do array addition.
I understand the iommu for the device (if correctly set and initialized)
will do the (io virtual address) -> (device bus address) translation because
the iommu subsystem would have setup the page table for the MMU(for kernel
virtual address) and IOMMU(for io virtual address) for the buffers so that
they can point to the same physical(or bus) address. I'm almost sure this
will work but can't do a test now. Is my understanding correct? can someone
well versed in the linux driver confirm me this approach is correct?

A question about iommu, passing to the device user data containing pointers

Chan Kim