Related
I'm running Linux 5.1 on a Cyclone V SoC, which is an FPGA with two ARMv7 cores in one chip. My goal is to gather lots of data from an external interface and stream (part of) this data out through a TCP socket. The challenge here is that the data rate is very high and could come close to saturating the GbE interface. I have a working implementation that just uses write() calls to the socket, but it tops out at 55MB/s; roughly half the theoretical GbE limit. I'm now trying to get zero-copy TCP transmission to work to increase the throughput, but I'm hitting a wall.
To get the data out of the FPGA into Linux user-space, I've written a kernel driver. This driver uses a DMA block in the FPGA to copy a large amount of data from an external interface into DDR3 memory attached to the ARMv7 cores. The driver allocates this memory as a bunch of contiguous 1MB buffers when probed using dma_alloc_coherent() with GFP_USER, and exposes these to the userspace application by implementing mmap() on a file in /dev/ and returning an address to the application using dma_mmap_coherent() on the preallocated buffers.
So far so good; the user-space application is seeing valid data and the throughput is more than enough at >360MB/s with room to spare (the external interface is not fast enough to really see what the upper bound is).
To implement zero-copy TCP networking, my first approach was to use SO_ZEROCOPY on the socket:
sent_bytes = send(fd, buf, len, MSG_ZEROCOPY);
if (sent_bytes < 0) {
perror("send");
return -1;
}
However, this results in send: Bad address.
After googling for a bit, my second approach was to use a pipe and splice() followed by vmsplice():
ssize_t sent_bytes;
int pipes[2];
struct iovec iov = {
.iov_base = buf,
.iov_len = len
};
pipe(pipes);
sent_bytes = vmsplice(pipes[1], &iov, 1, 0);
if (sent_bytes < 0) {
perror("vmsplice");
return -1;
}
sent_bytes = splice(pipes[0], 0, fd, 0, sent_bytes, SPLICE_F_MOVE);
if (sent_bytes < 0) {
perror("splice");
return -1;
}
However, the result is the same: vmsplice: Bad address.
Note that if I replace the call to vmsplice() or send() to a function that just prints the data pointed to by buf (or a send() without MSG_ZEROCOPY), everything is working just fine; so the data is accessible to userspace, but the vmsplice()/send(..., MSG_ZEROCOPY) calls seem unable to handle it.
What am I missing here? Is there any way of using zero-copy TCP sending with a user-space address obtained from a kernel driver through dma_mmap_coherent()? Is there another approach I could use?
UPDATE
So I dove a bit deeper into the sendmsg() MSG_ZEROCOPY path in the kernel, and the call that eventually fails is get_user_pages_fast(). This call returns -EFAULT because check_vma_flags() finds the VM_PFNMAP flag set in the vma. This flag is apparently set when the pages are mapped into user space using remap_pfn_range() or dma_mmap_coherent(). My next approach is to find another way to mmap these pages.
As I posted in an update in my question, the underlying problem is that zerocopy networking does not work for memory that has been mapped using remap_pfn_range() (which dma_mmap_coherent() happens to use under the hood as well). The reason is that this type of memory (with the VM_PFNMAP flag set) does not have metadata in the form of struct page* associated with each page, which it needs.
The solution then is to allocate the memory in a way that struct page*s are associated with the memory.
The workflow that now works for me to allocate the memory is:
Use struct page* page = alloc_pages(GFP_USER, page_order); to allocate a block of contiguous physical memory, where the number of contiguous pages that will be allocated is given by 2**page_order.
Split the high-order/compound page into 0-order pages by calling split_page(page, page_order);. This now means that struct page* page has become an array with 2**page_order entries.
Now to submit such a region to the DMA (for data reception):
dma_addr = dma_map_page(dev, page, 0, length, DMA_FROM_DEVICE);
dma_desc = dmaengine_prep_slave_single(dma_chan, dma_addr, length, DMA_DEV_TO_MEM, 0);
dmaengine_submit(dma_desc);
When we get a callback from the DMA that the transfer has finished, we need to unmap the region to transfer ownership of this block of memory back to the CPU, which takes care of caches to make sure we're not reading stale data:
dma_unmap_page(dev, dma_addr, length, DMA_FROM_DEVICE);
Now, when we want to implement mmap(), all we really have to do is call vm_insert_page() repeatedly for all of the 0-order pages that we pre-allocated:
static int my_mmap(struct file *file, struct vm_area_struct *vma) {
int res;
...
for (i = 0; i < 2**page_order; ++i) {
if ((res = vm_insert_page(vma, vma->vm_start + i*PAGE_SIZE, &page[i])) < 0) {
break;
}
}
vma->vm_flags |= VM_LOCKED | VM_DONTCOPY | VM_DONTEXPAND | VM_DENYWRITE;
...
return res;
}
When the file is closed, don't forget to free the pages:
for (i = 0; i < 2**page_order; ++i) {
__free_page(&dev->shm[i].pages[i]);
}
Implementing mmap() this way now allows a socket to use this buffer for sendmsg() with the MSG_ZEROCOPY flag.
Although this works, there are two things that don't sit well with me with this approach:
You can only allocate power-of-2-sized buffers with this method, although you could implement logic to call alloc_pages as many times as needed with decreasing orders to get any size buffer made up of sub-buffers of varying sizes. This will then require some logic to tie these buffers together in the mmap() and to DMA them with scatter-gather (sg) calls rather than single.
split_page() says in its documentation:
* Note: this is probably too low level an operation for use in drivers.
* Please consult with lkml before using this in your driver.
These issues would be easily solved if there was some interface in the kernel to allocate an arbitrary amount of contiguous physical pages. I don't know why there isn't, but I don't find the above issues so important as to go digging into why this isn't available / how to implement it :-)
Maybe this will help you to understand why alloc_pages requires a power-of-2 page number.
To optimize the page allocation process(and decrease external fragmentations), which is frequently engaged, Linux kernel developed per-cpu page cache and buddy-allocator to allocate memory(there is another allocator, slab, to serve memory allocations that are smaller than a page).
Per-cpu page cache serve the one-page allocation request, while buddy-allocator keeps 11 lists, each containing 2^{0-10} physical pages respectively. These lists perform well when allocate and free pages, and of course, the premise is you are requesting a power-of-2-sized buffer.
My objective is to be able to determine how many bytes have been transferred into the write end of a pipe. Perhaps, one would need to access the f_pos member of the struct file structure from linux/fs.h associated with this pipe.
struct file snipfrom fs.h
Is it possible to access this value from a userspace program? Again, I'd just like to be able to determine (perhaps based on the f_pos value) how many bytes are stored in the kernel buffer backing the pipe.
I have a feeling this isn't possible and one has to keep reading until read(int fd, void *buf, size_t count) returns less bytes than count.. then at this point, all bytes have been "emptied out" I assume..
Amount of bytes available for read from the pipe can be requested by
ioctl(fd, FIONREAD, &nbytes);
Here fd is a file descriptor, and variable nbytes, where result will be stored, is int variable.
Taken from: man 7 pipe.
Amount of bytes available for write is a different story.
I am writing an application server that processes images (large data). I am trying to minimize copies when sending image data back to clients. The processed images I need to send to clients are in buffers obtained from jemalloc. The ways I have thought of sending the data back to the client is:
1) Simple write call.
// Allocate buffer buf.
// Store image data in this buffer.
write(socket, buf, len);
2) I obtain the buffer through mmap instead of jemalloc, though I presume jemalloc already creates the buffer using mmap. I then make a simple call to write.
buf = mmap(file, len); // Imagine proper options.
// Store image data in this buffer.
write(socket, buf, len);
3) I obtain a buffer through mmap like before. I then use sendfile to send the data:
buf = mmap(in_fd, len); // Imagine proper options.
// Store image data in this buffer.
int rc;
rc = sendfile(out_fd, file, &offset, count);
// Deal with rc.
It seems like (1) and (2) will probably do the same thing given jemalloc probably allocates memory through mmap in the first place. I am not sure about (3) though. Will this really lead to any benefits? Figure 4 on this article on Linux zero-copy methods suggests that a further copy can be prevented using sendfile:
no data is copied into the socket buffer. Instead, only descriptors
with information about the whereabouts and length of the data are
appended to the socket buffer. The DMA engine passes data directly
from the kernel buffer to the protocol engine, thus eliminating the
remaining final copy.
This seems like a win if everything works out. I don't know if my mmaped buffer counts as a kernel buffer though. Also I don't know when it is safe to re-use this buffer. Since the fd and length is the only thing appended to the socket buffer, I assume that the kernel actually writes this data to the socket asynchronously. If it does what does the return from sendfile signify? How would I know when to re-use this buffer?
So my questions are:
What is the fastest way to write large buffers (images in my case) to a socket? The images are held in memory.
Is it a good idea to call sendfile on a mmapped file? If yes, what are the gotchas? Does this even lead to any wins?
It seems like my suspicions were correct. I got my information from this article. Quoting from it:
Also these network write system calls, including sendfile, might and
in many cases do return before the data sent over TCP by the method
call has been acknowledged. These methods return as soon as all data
is written into the socket buffers (sk buff) and is pushed to the TCP
write queue, the TCP engine can manage alone from that point on. In
other words at the time sendfile returns the last TCP send window is
not actually sent to the remote host but queued. In cases where
scatter-gather DMA is supported there is no seperate buffer which
holds these bytes, rather the buffers(sk buffs) just hold pointers to
the pages of OS buffer cache, where the contents of file is located.
This might lead to a race condition if we modify the content of the
file corresponding to the data in the last TCP send window as soon as
sendfile is returned. As a result TCP engine may send newly written
data to the remote host instead of what we originally intended to
send.
Provided the buffer from a mmapped file is even considered "DMA-able", seems like there is no way to know when it is safe to re-use it without an explicit acknowledgement (over the network) from the actual client. I might have to stick to simple write calls and incur the extra copy. There is a paper (also from the article) with more details.
Edit: This article on the splice call also shows the problems. Quoting it:
Be aware, when splicing data from a mmap'ed buffer to a network
socket, it is not possible to say when all data has been sent. Even if
splice() returns, the network stack may not have sent all data yet. So
reusing the buffer may overwrite unsent data.
For cases 1 and 2 - does the operation you marked as // Store image data in this buffer require any conversion? Is it just plain copy from the memory to buf?
If it's just plain copy, you can use write directly on the pointer obtained from jemalloc.
Assuming that img is a pointer obtained from jemalloc and size is a size of your image, just run following code:
int result;
int sent=0;
while(sent<size) {
result=write(socket,img+sent,size-sent);
if(result<0) {
/* error handling here */
break;
}
sent+=result;
}
It is working correctly for blocking I/O (the default behavior). If you need to write a data in a non-blocking manner, you should be able to rework the code on your own, but now you have the idea.
For case 3 - sendfile is for sending data from one descriptor to another. That means you can, for example, send data from file directly to tcp socket and you don't need to allocate any additional buffer. So, if the image you want to send to a client is in a file, just go for a sendfile. If you have it in memory (because you processed it somehow, or just generated), use the approach I mentioned earlier.
In scatter and gather (i.e. readv and writev), Linux reads into multiple buffers and writes from multiple buffers.
If say, I have a vector of 3 buffers, I can use readv, OR I can use a single buffer, which is of combined size of 3 buffers and do fread.
Hence, I am confused: For which cases should scatter/gather be used and when should a single large buffer be used?
The main convenience offered by readv, writev is:
It allows working with non contiguous blocks of data. i.e. buffers need not be part of an array, but separately allocated.
The I/O is 'atomic'. i.e. If you do a writev, all the elements in the vector will be written in one contiguous operation, and writes done by other processes will not occur in between them.
e.g. say, your data is naturally segmented, and comes from different sources:
struct foo *my_foo;
struct bar *my_bar;
struct baz *my_baz;
my_foo = get_my_foo();
my_bar = get_my_bar();
my_baz = get_my_baz();
Now, all three 'buffers' are not one big contiguous block. But you want to write them contiguously into a file, for whatever reason (say for example, they are fields in a file header for a file format).
If you use write you have to choose between:
Copying them over into one block of memory using, say, memcpy (overhead), followed by a single write call. Then the write will be atomic.
Making three separate calls to write (overhead). Also, write calls from other processes can intersperse between these writes (not atomic).
If you use writev instead, its all good:
You make exactly one system call, and no memcpy to make a single buffer from the three.
Also, the three buffers are written atomically, as one block write. i.e. if other processes also write, then these writes will not come in between the writes of the three vectors.
So you would do something like:
struct iovec iov[3];
iov[0].iov_base = my_foo;
iov[0].iov_len = sizeof (struct foo);
iov[1].iov_base = my_bar;
iov[1].iov_len = sizeof (struct bar);
iov[2].iov_base = my_baz;
iov[2].iov_len = sizeof (struct baz);
bytes_written = writev (fd, iov, 3);
Sources:
http://pubs.opengroup.org/onlinepubs/009604499/functions/writev.html
http://linux.die.net/man/2/readv
I want to get data from a DMA enabled, PCIe hardware device into user-space as quickly as possible.
Q: How do I combine "direct I/O to user-space with/and/via a DMA transfer"
Reading through LDD3, it seems that I need to perform a few different types of IO operations!?
dma_alloc_coherent gives me the physical address that I can pass to the hardware device.
But would need to have setup get_user_pages and perform a copy_to_user type call when the transfer completes. This seems a waste, asking the Device to DMA into kernel memory (acting as buffer) then transferring it again to user-space.
LDD3 p453: /* Only now is it safe to access the buffer, copy to user, etc. */
What I ideally want is some memory that:
I can use in user-space (Maybe request driver via a ioctl call to create DMA'able memory/buffer?)
I can get a physical address from to pass to the device so that all user-space has to do is perform a read on the driver
the read method would activate the DMA transfer, block waiting for the DMA complete interrupt and release the user-space read afterwards (user-space is now safe to use/read memory).
Do I need single-page streaming mappings, setup mapping and user-space buffers mapped with get_user_pages dma_map_page?
My code so far sets up get_user_pages at the given address from user-space (I call this the Direct I/O part). Then, dma_map_page with a page from get_user_pages. I give the device the return value from dma_map_page as the DMA physical transfer address.
I am using some kernel modules as reference: drivers_scsi_st.c and drivers-net-sh_eth.c. I would look at infiniband code, but cant find which one is the most basic!
Many thanks in advance.
I'm actually working on exactly the same thing right now and I'm going the ioctl() route. The general idea is for user space to allocate the buffer which will be used for the DMA transfer and an ioctl() will be used to pass the size and address of this buffer to the device driver. The driver will then use scatter-gather lists along with the streaming DMA API to transfer data directly to and from the device and user-space buffer.
The implementation strategy I'm using is that the ioctl() in the driver enters a loop that DMA's the userspace buffer in chunks of 256k (which is the hardware imposed limit for how many scatter/gather entries it can handle). This is isolated inside a function that blocks until each transfer is complete (see below). When all bytes are transfered or the incremental transfer function returns an error the ioctl() exits and returns to userspace
Pseudo code for the ioctl()
/*serialize all DMA transfers to/from the device*/
if (mutex_lock_interruptible( &device_ptr->mtx ) )
return -EINTR;
chunk_data = (unsigned long) user_space_addr;
while( *transferred < total_bytes && !ret ) {
chunk_bytes = total_bytes - *transferred;
if (chunk_bytes > HW_DMA_MAX)
chunk_bytes = HW_DMA_MAX; /* 256kb limit imposed by my device */
ret = transfer_chunk(device_ptr, chunk_data, chunk_bytes, transferred);
chunk_data += chunk_bytes;
chunk_offset += chunk_bytes;
}
mutex_unlock(&device_ptr->mtx);
Pseudo code for incremental transfer function:
/*Assuming the userspace pointer is passed as an unsigned long, */
/*calculate the first,last, and number of pages being transferred via*/
first_page = (udata & PAGE_MASK) >> PAGE_SHIFT;
last_page = ((udata+nbytes-1) & PAGE_MASK) >> PAGE_SHIFT;
first_page_offset = udata & PAGE_MASK;
npages = last_page - first_page + 1;
/* Ensure that all userspace pages are locked in memory for the */
/* duration of the DMA transfer */
down_read(¤t->mm->mmap_sem);
ret = get_user_pages(current,
current->mm,
udata,
npages,
is_writing_to_userspace,
0,
&pages_array,
NULL);
up_read(¤t->mm->mmap_sem);
/* Map a scatter-gather list to point at the userspace pages */
/*first*/
sg_set_page(&sglist[0], pages_array[0], PAGE_SIZE - fp_offset, fp_offset);
/*middle*/
for(i=1; i < npages-1; i++)
sg_set_page(&sglist[i], pages_array[i], PAGE_SIZE, 0);
/*last*/
if (npages > 1) {
sg_set_page(&sglist[npages-1], pages_array[npages-1],
nbytes - (PAGE_SIZE - fp_offset) - ((npages-2)*PAGE_SIZE), 0);
}
/* Do the hardware specific thing to give it the scatter-gather list
and tell it to start the DMA transfer */
/* Wait for the DMA transfer to complete */
ret = wait_event_interruptible_timeout( &device_ptr->dma_wait,
&device_ptr->flag_dma_done, HZ*2 );
if (ret == 0)
/* DMA operation timed out */
else if (ret == -ERESTARTSYS )
/* DMA operation interrupted by signal */
else {
/* DMA success */
*transferred += nbytes;
return 0;
}
The interrupt handler is exceptionally brief:
/* Do hardware specific thing to make the device happy */
/* Wake the thread waiting for this DMA operation to complete */
device_ptr->flag_dma_done = 1;
wake_up_interruptible(device_ptr->dma_wait);
Please note that this is just a general approach, I've been working on this driver for the last few weeks and have yet to actually test it... So please, don't treat this pseudo code as gospel and be sure to double check all logic and parameters ;-).
You basically have the right idea: in 2.1, you can just have userspace allocate any old memory. You do want it page-aligned, so posix_memalign() is a handy API to use.
Then have userspace pass in the userspace virtual address and size of this buffer somehow; ioctl() is a good quick and dirty way to do this. In the kernel, allocate an appropriately sized buffer array of struct page* -- user_buf_size/PAGE_SIZE entries -- and use get_user_pages() to get a list of struct page* for the userspace buffer.
Once you have that, you can allocate an array of struct scatterlist that is the same size as your page array and loop through the list of pages doing sg_set_page(). After the sg list is set up, you do dma_map_sg() on the array of scatterlist and then you can get the sg_dma_address and sg_dma_len for each entry in the scatterlist (note you have to use the return value of dma_map_sg() because you may end up with fewer mapped entries because things might get merged by the DMA mapping code).
That gives you all the bus addresses to pass to your device, and then you can trigger the DMA and wait for it however you want. The read()-based scheme you have is probably fine.
You can refer to drivers/infiniband/core/umem.c, specifically ib_umem_get(), for some code that builds up this mapping, although the generality that that code needs to deal with may make it a bit confusing.
Alternatively, if your device doesn't handle scatter/gather lists too well and you want contiguous memory, you could use get_free_pages() to allocate a physically contiguous buffer and use dma_map_page() on that. To give userspace access to that memory, your driver just needs to implement an mmap method instead of the ioctl as described above.
At some point I wanted to allow user-space application to allocate DMA buffers and get it mapped to user-space and get the physical address to be able to control my device and do DMA transactions (bus mastering) entirely from user-space, totally bypassing the Linux kernel. I have used a little bit different approach though. First I started with a minimal kernel module that was initializing/probing PCIe device and creating a character device. That driver then allowed a user-space application to do two things:
Map PCIe device's I/O bar into user-space using remap_pfn_range() function.
Allocate and free DMA buffers, map them to user space and pass a physical bus address to user-space application.
Basically, it boils down to a custom implementation of mmap() call (though file_operations). One for I/O bar is easy:
struct vm_operations_struct a2gx_bar_vma_ops = {
};
static int a2gx_cdev_mmap_bar2(struct file *filp, struct vm_area_struct *vma)
{
struct a2gx_dev *dev;
size_t size;
size = vma->vm_end - vma->vm_start;
if (size != 134217728)
return -EIO;
dev = filp->private_data;
vma->vm_ops = &a2gx_bar_vma_ops;
vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
vma->vm_private_data = dev;
if (remap_pfn_range(vma, vma->vm_start,
vmalloc_to_pfn(dev->bar2),
size, vma->vm_page_prot))
{
return -EAGAIN;
}
return 0;
}
And another one that allocates DMA buffers using pci_alloc_consistent() is a little bit more complicated:
static void a2gx_dma_vma_close(struct vm_area_struct *vma)
{
struct a2gx_dma_buf *buf;
struct a2gx_dev *dev;
buf = vma->vm_private_data;
dev = buf->priv_data;
pci_free_consistent(dev->pci_dev, buf->size, buf->cpu_addr, buf->dma_addr);
buf->cpu_addr = NULL; /* Mark this buffer data structure as unused/free */
}
struct vm_operations_struct a2gx_dma_vma_ops = {
.close = a2gx_dma_vma_close
};
static int a2gx_cdev_mmap_dma(struct file *filp, struct vm_area_struct *vma)
{
struct a2gx_dev *dev;
struct a2gx_dma_buf *buf;
size_t size;
unsigned int i;
/* Obtain a pointer to our device structure and calculate the size
of the requested DMA buffer */
dev = filp->private_data;
size = vma->vm_end - vma->vm_start;
if (size < sizeof(unsigned long))
return -EINVAL; /* Something fishy is happening */
/* Find a structure where we can store extra information about this
buffer to be able to release it later. */
for (i = 0; i < A2GX_DMA_BUF_MAX; ++i) {
buf = &dev->dma_buf[i];
if (buf->cpu_addr == NULL)
break;
}
if (buf->cpu_addr != NULL)
return -ENOBUFS; /* Oops, hit the limit of allowed number of
allocated buffers. Change A2GX_DMA_BUF_MAX and
recompile? */
/* Allocate consistent memory that can be used for DMA transactions */
buf->cpu_addr = pci_alloc_consistent(dev->pci_dev, size, &buf->dma_addr);
if (buf->cpu_addr == NULL)
return -ENOMEM; /* Out of juice */
/* There is no way to pass extra information to the user. And I am too lazy
to implement this mmap() call using ioctl(). So we simply tell the user
the bus address of this buffer by copying it to the allocated buffer
itself. Hacks, hacks everywhere. */
memcpy(buf->cpu_addr, &buf->dma_addr, sizeof(buf->dma_addr));
buf->size = size;
buf->priv_data = dev;
vma->vm_ops = &a2gx_dma_vma_ops;
vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
vma->vm_private_data = buf;
/*
* Map this DMA buffer into user space.
*/
if (remap_pfn_range(vma, vma->vm_start,
vmalloc_to_pfn(buf->cpu_addr),
size, vma->vm_page_prot))
{
/* Out of luck, rollback... */
pci_free_consistent(dev->pci_dev, buf->size, buf->cpu_addr,
buf->dma_addr);
buf->cpu_addr = NULL;
return -EAGAIN;
}
return 0; /* All good! */
}
Once those are in place, user space application can pretty much do everything — control the device by reading/writing from/to I/O registers, allocate and free DMA buffers of arbitrary size, and have the device perform DMA transactions. The only missing part is interrupt-handling. I was doing polling in user space, burning my CPU, and had interrupts disabled.
Hope it helps. Good Luck!
I'm getting confused with the direction to implement. I want to...
Consider the application when designing a driver.
What is the nature of data movement, frequency, size and what else might be going on in the system?
Is the traditional read/write API sufficient?
Is direct mapping the device into user space OK?
Is a reflective (semi-coherent) shared memory desirable?
Manually manipulating data (read/write) is a pretty good option if the data lends itself to being well understood. Using general purpose VM and read/write may be sufficient with an inline copy. Direct mapping non cachable accesses to the peripheral is convenient, but can be clumsy. If the access is the relatively infrequent movement of large blocks, it may make sense to use regular memory, have the drive pin, translate addresses, DMA and release the pages. As an optimization, the pages (maybe huge) can be pre pinned and translated; the drive then can recognize the prepared memory and avoid the complexities of dynamic translation. If there are lots of little I/O operations, having the drive run asynchronously makes sense. If elegance is important, the VM dirty page flag can be used to automatically identify what needs to be moved and a (meta_sync()) call can be used to flush pages. Perhaps a mixture of the above works...
Too often people don't look at the larger problem, before digging into the details. Often the simplest solutions are sufficient. A little effort constructing a behavioral model can help guide what API is preferable.
first_page_offset = udata & PAGE_MASK;
It seems wrong. It should be either:
first_page_offset = udata & ~PAGE_MASK;
or
first_page_offset = udata & (PAGE_SIZE - 1)
It is worth mention that driver with Scatter-Gather DMA support and user space memory allocation is most efficient and has highest performance. However in case we don't need high performance or we want to develop a driver in some simplified conditions we can use some tricks.
Give up zero copy design. It is worth to consider when data throughput is not too big. In such a design data can by copied to user by
copy_to_user(user_buffer, kernel_dma_buffer, count);
user_buffer might be for example buffer argument in character device read() system call implementation. We still need to take care of kernel_dma_buffer allocation. It might by memory obtained from dma_alloc_coherent() call for example.
The another trick is to limit system memory at the boot time and then use it as huge contiguous DMA buffer. It is especially useful during driver and FPGA DMA controller development and rather not recommended in production environments. Lets say PC has 32GB of RAM. If we add mem=20GB to kernel boot parameters list we can use 12GB as huge contiguous dma buffer. To map this memory to user space simply implement mmap() as
remap_pfn_range(vma,
vma->vm_start,
(0x500000000 >> PAGE_SHIFT) + vma->vm_pgoff,
vma->vm_end - vma->vm_start,
vma->vm_page_prot)
Of course this 12GB is completely omitted by OS and can be used only by process which has mapped it into its address space. We can try to avoid it by using Contiguous Memory Allocator (CMA).
Again above tricks will not replace full Scatter-Gather, zero copy DMA driver, but are useful during development time or in some less performance platforms.