Device drivers have to be careful when using memory. As they are part of the Linux kernel they cannot use virtual memory. Each time they run, maybe as an interrupt is received or as a bottom half or task queue handler is scheduled, the current process may change. The device driver cannot rely on a particular process running even if it is doing work on its behalf. Like the rest of the kernel, device drivers use data structures to keep track of the device that it is controlling. These data structures can be statically allocated, part of the device driver's code, but that would be wasteful as it makes the kernel larger than it need be. Most device drivers allocate kernel, non-paged, memory to hold their data.
Linux provides kernel memory allocation and deallocation routines and it is these that the device drivers use. Kernel memory is allocated in chunks that are powers of 2. For example 128 or 512 bytes, even if the device driver asks for less. The number of bytes that the device driver requests is rounded up to the next block size boundry. This makes kernel memory deallocation easier as the smaller free blocks can be recombined into bigger blocks.
It may be that Linux needs to do quite a lot of extra work when the kernel memory is requested. If the amount of free memory is low, physical pages may need to be discarded or written to the swap device. Normally, Linux would suspend the requestor, putting the process onto a wait queue until there is enough physical memory. Not all device drivers (or indeed Linux kernel code) may want this to happen and so the kernel memory allocation routines can be requested to fail if they cannot immediately allocate memory. If the device driver wishes to DMA to or from the allocated memory it can also specify that the memory is DMA'able. This way it is the Linux kernel that needs to understand what constitutes DMA'able memory for this system, and not the device driver.