linuxmodule: fix md5_append wrt non-virt data

[pallas]

sg_init_one uses sg_init_buf, which only works with kmalloc'd data. In particular, this excludes anything vmalloc'd as well as anything in the rodata section of the module. The code that was failing for me was essentially

    String str("data");
    uint8_t digest[16];

    md5_state_t state;
    md5_init(&state);
    md5_append(&state, str.data(), str.length());
    md5_finish(&state, digest);
    md5_free(&state);

and the work-around was to invoke String::mutable_data() prior to the md5_append, which kmalloc's and makes String::data() safe for sg_init_buf.

The better solution is to call virt_addr_valid, only using sg_init_one if true; otherwise, we must use vmalloc_to_page on individual pages, which will walk the page-table and should do the Right Thing.

Since vmalloc'd data might not be continuous, we must iterate over the individual pages that make up the allocation. Because this is expensive to begin with, we call update for each page using a single scatterlist rather than allocating an sg_table and calling update once. This also simplifies the code.

I tested this by adding

    click_chatter("%p %d %d %d", data, off, len, nbytes);

inside the loop and watching it "do the Right Thing" to a large vmalloc allocation, verifying that the hash was correct using md5sum.

[kohler]

Epic. Thanks. (Maybe we should've just used our own MD5.)