blob: a0b6250add79c9c144fd1185fa91314febaa4241 [file] [log] [blame]
Dynamic DMA mapping Guide
David S. Miller <>
Richard Henderson <>
Jakub Jelinek <>
This is a guide to device driver writers on how to use the DMA API
with example pseudo-code. For a concise description of the API, see
Most of the 64bit platforms have special hardware that translates bus
addresses (DMA addresses) into physical addresses. This is similar to
how page tables and/or a TLB translates virtual addresses to physical
addresses on a CPU. This is needed so that e.g. PCI devices can
access with a Single Address Cycle (32bit DMA address) any page in the
64bit physical address space. Previously in Linux those 64bit
platforms had to set artificial limits on the maximum RAM size in the
system, so that the virt_to_bus() static scheme works (the DMA address
translation tables were simply filled on bootup to map each bus
address to the physical page __pa(bus_to_virt())).
So that Linux can use the dynamic DMA mapping, it needs some help from the
drivers, namely it has to take into account that DMA addresses should be
mapped only for the time they are actually used and unmapped after the DMA
The following API will work of course even on platforms where no such
hardware exists.
Note that the DMA API works with any bus independent of the underlying
microprocessor architecture. You should use the DMA API rather than
the bus specific DMA API (e.g. pci_dma_*).
First of all, you should make sure
#include <linux/dma-mapping.h>
is in your driver. This file will obtain for you the definition of the
dma_addr_t (which can hold any valid DMA address for the platform)
type which should be used everywhere you hold a DMA (bus) address
returned from the DMA mapping functions.
What memory is DMA'able?
The first piece of information you must know is what kernel memory can
be used with the DMA mapping facilities. There has been an unwritten
set of rules regarding this, and this text is an attempt to finally
write them down.
If you acquired your memory via the page allocator
(i.e. __get_free_page*()) or the generic memory allocators
(i.e. kmalloc() or kmem_cache_alloc()) then you may DMA to/from
that memory using the addresses returned from those routines.
This means specifically that you may _not_ use the memory/addresses
returned from vmalloc() for DMA. It is possible to DMA to the
_underlying_ memory mapped into a vmalloc() area, but this requires
walking page tables to get the physical addresses, and then
translating each of those pages back to a kernel address using
something like __va(). [ EDIT: Update this when we integrate
Gerd Knorr's generic code which does this. ]
This rule also means that you may use neither kernel image addresses
(items in data/text/bss segments), nor module image addresses, nor
stack addresses for DMA. These could all be mapped somewhere entirely
different than the rest of physical memory. Even if those classes of
memory could physically work with DMA, you'd need to ensure the I/O
buffers were cacheline-aligned. Without that, you'd see cacheline
sharing problems (data corruption) on CPUs with DMA-incoherent caches.
(The CPU could write to one word, DMA would write to a different one
in the same cache line, and one of them could be overwritten.)
Also, this means that you cannot take the return of a kmap()
call and DMA to/from that. This is similar to vmalloc().
What about block I/O and networking buffers? The block I/O and
networking subsystems make sure that the buffers they use are valid
for you to DMA from/to.
DMA addressing limitations
Does your device have any DMA addressing limitations? For example, is
your device only capable of driving the low order 24-bits of address?
If so, you need to inform the kernel of this fact.
By default, the kernel assumes that your device can address the full
32-bits. For a 64-bit capable device, this needs to be increased.
And for a device with limitations, as discussed in the previous
paragraph, it needs to be decreased.
Special note about PCI: PCI-X specification requires PCI-X devices to
support 64-bit addressing (DAC) for all transactions. And at least
one platform (SGI SN2) requires 64-bit consistent allocations to
operate correctly when the IO bus is in PCI-X mode.
For correct operation, you must interrogate the kernel in your device
probe routine to see if the DMA controller on the machine can properly
support the DMA addressing limitation your device has. It is good
style to do this even if your device holds the default setting,
because this shows that you did think about these issues wrt. your
The query is performed via a call to dma_set_mask():
int dma_set_mask(struct device *dev, u64 mask);
The query for consistent allocations is performed via a call to
int dma_set_coherent_mask(struct device *dev, u64 mask);
Here, dev is a pointer to the device struct of your device, and mask
is a bit mask describing which bits of an address your device
supports. It returns zero if your card can perform DMA properly on
the machine given the address mask you provided. In general, the
device struct of your device is embedded in the bus specific device
struct of your device. For example, a pointer to the device struct of
your PCI device is pdev->dev (pdev is a pointer to the PCI device
struct of your device).
If it returns non-zero, your device cannot perform DMA properly on
this platform, and attempting to do so will result in undefined
behavior. You must either use a different mask, or not use DMA.
This means that in the failure case, you have three options:
1) Use another DMA mask, if possible (see below).
2) Use some non-DMA mode for data transfer, if possible.
3) Ignore this device and do not initialize it.
It is recommended that your driver print a kernel KERN_WARNING message
when you end up performing either #2 or #3. In this manner, if a user
of your driver reports that performance is bad or that the device is not
even detected, you can ask them for the kernel messages to find out
exactly why.
The standard 32-bit addressing device would do something like this:
if (dma_set_mask(dev, DMA_BIT_MASK(32))) {
"mydev: No suitable DMA available.\n");
goto ignore_this_device;
Another common scenario is a 64-bit capable device. The approach here
is to try for 64-bit addressing, but back down to a 32-bit mask that
should not fail. The kernel may fail the 64-bit mask not because the
platform is not capable of 64-bit addressing. Rather, it may fail in
this case simply because 32-bit addressing is done more efficiently
than 64-bit addressing. For example, Sparc64 PCI SAC addressing is
more efficient than DAC addressing.
Here is how you would handle a 64-bit capable device which can drive
all 64-bits when accessing streaming DMA:
int using_dac;
if (!dma_set_mask(dev, DMA_BIT_MASK(64))) {
using_dac = 1;
} else if (!dma_set_mask(dev, DMA_BIT_MASK(32))) {
using_dac = 0;
} else {
"mydev: No suitable DMA available.\n");
goto ignore_this_device;
If a card is capable of using 64-bit consistent allocations as well,
the case would look like this:
int using_dac, consistent_using_dac;
if (!dma_set_mask(dev, DMA_BIT_MASK(64))) {
using_dac = 1;
consistent_using_dac = 1;
dma_set_coherent_mask(dev, DMA_BIT_MASK(64));
} else if (!dma_set_mask(dev, DMA_BIT_MASK(32))) {
using_dac = 0;
consistent_using_dac = 0;
dma_set_coherent_mask(dev, DMA_BIT_MASK(32));
} else {
"mydev: No suitable DMA available.\n");
goto ignore_this_device;
dma_set_coherent_mask() will always be able to set the same or a
smaller mask as dma_set_mask(). However for the rare case that a
device driver only uses consistent allocations, one would have to
check the return value from dma_set_coherent_mask().
Finally, if your device can only drive the low 24-bits of
address you might do something like:
if (dma_set_mask(dev, DMA_BIT_MASK(24))) {
"mydev: 24-bit DMA addressing not available.\n");
goto ignore_this_device;
When dma_set_mask() is successful, and returns zero, the kernel saves
away this mask you have provided. The kernel will use this
information later when you make DMA mappings.
There is a case which we are aware of at this time, which is worth
mentioning in this documentation. If your device supports multiple
functions (for example a sound card provides playback and record
functions) and the various different functions have _different_
DMA addressing limitations, you may wish to probe each mask and
only provide the functionality which the machine can handle. It
is important that the last call to dma_set_mask() be for the
most specific mask.
Here is pseudo-code showing how this might be done:
struct my_sound_card *card;
struct device *dev;
if (!dma_set_mask(dev, PLAYBACK_ADDRESS_BITS)) {
card->playback_enabled = 1;
} else {
card->playback_enabled = 0;
printk(KERN_WARNING "%s: Playback disabled due to DMA limitations.\n",
if (!dma_set_mask(dev, RECORD_ADDRESS_BITS)) {
card->record_enabled = 1;
} else {
card->record_enabled = 0;
printk(KERN_WARNING "%s: Record disabled due to DMA limitations.\n",
A sound card was used as an example here because this genre of PCI
devices seems to be littered with ISA chips given a PCI front end,
and thus retaining the 16MB DMA addressing limitations of ISA.
Types of DMA mappings
There are two types of DMA mappings:
- Consistent DMA mappings which are usually mapped at driver
initialization, unmapped at the end and for which the hardware should
guarantee that the device and the CPU can access the data
in parallel and will see updates made by each other without any
explicit software flushing.
Think of "consistent" as "synchronous" or "coherent".
The current default is to return consistent memory in the low 32
bits of the bus space. However, for future compatibility you should
set the consistent mask even if this default is fine for your
Good examples of what to use consistent mappings for are:
- Network card DMA ring descriptors.
- SCSI adapter mailbox command data structures.
- Device firmware microcode executed out of
main memory.
The invariant these examples all require is that any CPU store
to memory is immediately visible to the device, and vice
versa. Consistent mappings guarantee this.
IMPORTANT: Consistent DMA memory does not preclude the usage of
proper memory barriers. The CPU may reorder stores to
consistent memory just as it may normal memory. Example:
if it is important for the device to see the first word
of a descriptor updated before the second, you must do
something like:
desc->word0 = address;
desc->word1 = DESC_VALID;
in order to get correct behavior on all platforms.
Also, on some platforms your driver may need to flush CPU write
buffers in much the same way as it needs to flush write buffers
found in PCI bridges (such as by reading a register's value
after writing it).
- Streaming DMA mappings which are usually mapped for one DMA
transfer, unmapped right after it (unless you use dma_sync_* below)
and for which hardware can optimize for sequential accesses.
This of "streaming" as "asynchronous" or "outside the coherency
Good examples of what to use streaming mappings for are:
- Networking buffers transmitted/received by a device.
- Filesystem buffers written/read by a SCSI device.
The interfaces for using this type of mapping were designed in
such a way that an implementation can make whatever performance
optimizations the hardware allows. To this end, when using
such mappings you must be explicit about what you want to happen.
Neither type of DMA mapping has alignment restrictions that come from
the underlying bus, although some devices may have such restrictions.
Also, systems with caches that aren't DMA-coherent will work better
when the underlying buffers don't share cache lines with other data.
Using Consistent DMA mappings.
To allocate and map large (PAGE_SIZE or so) consistent DMA regions,
you should do:
dma_addr_t dma_handle;
cpu_addr = dma_alloc_coherent(dev, size, &dma_handle, gfp);
where device is a struct device *. This may be called in interrupt
context with the GFP_ATOMIC flag.
Size is the length of the region you want to allocate, in bytes.
This routine will allocate RAM for that region, so it acts similarly to
__get_free_pages (but takes size instead of a page order). If your
driver needs regions sized smaller than a page, you may prefer using
the dma_pool interface, described below.
The consistent DMA mapping interfaces, for non-NULL dev, will by
default return a DMA address which is 32-bit addressable. Even if the
device indicates (via DMA mask) that it may address the upper 32-bits,
consistent allocation will only return > 32-bit addresses for DMA if
the consistent DMA mask has been explicitly changed via
dma_set_coherent_mask(). This is true of the dma_pool interface as
dma_alloc_coherent returns two values: the virtual address which you
can use to access it from the CPU and dma_handle which you pass to the
The cpu return address and the DMA bus master address are both
guaranteed to be aligned to the smallest PAGE_SIZE order which
is greater than or equal to the requested size. This invariant
exists (for example) to guarantee that if you allocate a chunk
which is smaller than or equal to 64 kilobytes, the extent of the
buffer you receive will not cross a 64K boundary.
To unmap and free such a DMA region, you call:
dma_free_coherent(dev, size, cpu_addr, dma_handle);
where dev, size are the same as in the above call and cpu_addr and
dma_handle are the values dma_alloc_coherent returned to you.
This function may not be called in interrupt context.
If your driver needs lots of smaller memory regions, you can write
custom code to subdivide pages returned by dma_alloc_coherent,
or you can use the dma_pool API to do that. A dma_pool is like
a kmem_cache, but it uses dma_alloc_coherent not __get_free_pages.
Also, it understands common hardware constraints for alignment,
like queue heads needing to be aligned on N byte boundaries.
Create a dma_pool like this:
struct dma_pool *pool;
pool = dma_pool_create(name, dev, size, align, alloc);
The "name" is for diagnostics (like a kmem_cache name); dev and size
are as above. The device's hardware alignment requirement for this
type of data is "align" (which is expressed in bytes, and must be a
power of two). If your device has no boundary crossing restrictions,
pass 0 for alloc; passing 4096 says memory allocated from this pool
must not cross 4KByte boundaries (but at that time it may be better to
go for dma_alloc_coherent directly instead).
Allocate memory from a dma pool like this:
cpu_addr = dma_pool_alloc(pool, flags, &dma_handle);
flags are SLAB_KERNEL if blocking is permitted (not in_interrupt nor
holding SMP locks), SLAB_ATOMIC otherwise. Like dma_alloc_coherent,
this returns two values, cpu_addr and dma_handle.
Free memory that was allocated from a dma_pool like this:
dma_pool_free(pool, cpu_addr, dma_handle);
where pool is what you passed to dma_pool_alloc, and cpu_addr and
dma_handle are the values dma_pool_alloc returned. This function
may be called in interrupt context.
Destroy a dma_pool by calling:
Make sure you've called dma_pool_free for all memory allocated
from a pool before you destroy the pool. This function may not
be called in interrupt context.
DMA Direction
The interfaces described in subsequent portions of this document
take a DMA direction argument, which is an integer and takes on
one of the following values:
One should provide the exact DMA direction if you know it.
DMA_TO_DEVICE means "from main memory to the device"
DMA_FROM_DEVICE means "from the device to main memory"
It is the direction in which the data moves during the DMA
You are _strongly_ encouraged to specify this as precisely
as you possibly can.
If you absolutely cannot know the direction of the DMA transfer,
specify DMA_BIDIRECTIONAL. It means that the DMA can go in
either direction. The platform guarantees that you may legally
specify this, and that it will work, but this may be at the
cost of performance for example.
The value DMA_NONE is to be used for debugging. One can
hold this in a data structure before you come to know the
precise direction, and this will help catch cases where your
direction tracking logic has failed to set things up properly.
Another advantage of specifying this value precisely (outside of
potential platform-specific optimizations of such) is for debugging.
Some platforms actually have a write permission boolean which DMA
mappings can be marked with, much like page protections in the user
program address space. Such platforms can and do report errors in the
kernel logs when the DMA controller hardware detects violation of the
permission setting.
Only streaming mappings specify a direction, consistent mappings
implicitly have a direction attribute setting of
The SCSI subsystem tells you the direction to use in the
'sc_data_direction' member of the SCSI command your driver is
working on.
For Networking drivers, it's a rather simple affair. For transmit
packets, map/unmap them with the DMA_TO_DEVICE direction
specifier. For receive packets, just the opposite, map/unmap them
with the DMA_FROM_DEVICE direction specifier.
Using Streaming DMA mappings
The streaming DMA mapping routines can be called from interrupt
context. There are two versions of each map/unmap, one which will
map/unmap a single memory region, and one which will map/unmap a
To map a single region, you do:
struct device *dev = &my_dev->dev;
dma_addr_t dma_handle;
void *addr = buffer->ptr;
size_t size = buffer->len;
dma_handle = dma_map_single(dev, addr, size, direction);
and to unmap it:
dma_unmap_single(dev, dma_handle, size, direction);
You should call dma_unmap_single when the DMA activity is finished, e.g.
from the interrupt which told you that the DMA transfer is done.
Using cpu pointers like this for single mappings has a disadvantage,
you cannot reference HIGHMEM memory in this way. Thus, there is a
map/unmap interface pair akin to dma_{map,unmap}_single. These
interfaces deal with page/offset pairs instead of cpu pointers.
struct device *dev = &my_dev->dev;
dma_addr_t dma_handle;
struct page *page = buffer->page;
unsigned long offset = buffer->offset;
size_t size = buffer->len;
dma_handle = dma_map_page(dev, page, offset, size, direction);
dma_unmap_page(dev, dma_handle, size, direction);
Here, "offset" means byte offset within the given page.
With scatterlists, you map a region gathered from several regions by:
int i, count = dma_map_sg(dev, sglist, nents, direction);
struct scatterlist *sg;
for_each_sg(sglist, sg, count, i) {
hw_address[i] = sg_dma_address(sg);
hw_len[i] = sg_dma_len(sg);
where nents is the number of entries in the sglist.
The implementation is free to merge several consecutive sglist entries
into one (e.g. if DMA mapping is done with PAGE_SIZE granularity, any
consecutive sglist entries can be merged into one provided the first one
ends and the second one starts on a page boundary - in fact this is a huge
advantage for cards which either cannot do scatter-gather or have very
limited number of scatter-gather entries) and returns the actual number
of sg entries it mapped them to. On failure 0 is returned.
Then you should loop count times (note: this can be less than nents times)
and use sg_dma_address() and sg_dma_len() macros where you previously
accessed sg->address and sg->length as shown above.
To unmap a scatterlist, just call:
dma_unmap_sg(dev, sglist, nents, direction);
Again, make sure DMA activity has already finished.
PLEASE NOTE: The 'nents' argument to the dma_unmap_sg call must be
the _same_ one you passed into the dma_map_sg call,
it should _NOT_ be the 'count' value _returned_ from the
dma_map_sg call.
Every dma_map_{single,sg} call should have its dma_unmap_{single,sg}
counterpart, because the bus address space is a shared resource (although
in some ports the mapping is per each BUS so less devices contend for the
same bus address space) and you could render the machine unusable by eating
all bus addresses.
If you need to use the same streaming DMA region multiple times and touch
the data in between the DMA transfers, the buffer needs to be synced
properly in order for the cpu and device to see the most uptodate and
correct copy of the DMA buffer.
So, firstly, just map it with dma_map_{single,sg}, and after each DMA
transfer call either:
dma_sync_single_for_cpu(dev, dma_handle, size, direction);
dma_sync_sg_for_cpu(dev, sglist, nents, direction);
as appropriate.
Then, if you wish to let the device get at the DMA area again,
finish accessing the data with the cpu, and then before actually
giving the buffer to the hardware call either:
dma_sync_single_for_device(dev, dma_handle, size, direction);
dma_sync_sg_for_device(dev, sglist, nents, direction);
as appropriate.
After the last DMA transfer call one of the DMA unmap routines
dma_unmap_{single,sg}. If you don't touch the data from the first dma_map_*
call till dma_unmap_*, then you don't have to call the dma_sync_*
routines at all.
Here is pseudo code which shows a situation in which you would need
to use the dma_sync_*() interfaces.
my_card_setup_receive_buffer(struct my_card *cp, char *buffer, int len)
dma_addr_t mapping;
mapping = dma_map_single(cp->dev, buffer, len, DMA_FROM_DEVICE);
cp->rx_buf = buffer;
cp->rx_len = len;
cp->rx_dma = mapping;
my_card_interrupt_handler(int irq, void *devid, struct pt_regs *regs)
struct my_card *cp = devid;
if (read_card_status(cp) == RX_BUF_TRANSFERRED) {
struct my_card_header *hp;
/* Examine the header to see if we wish
* to accept the data. But synchronize
* the DMA transfer with the CPU first
* so that we see updated contents.
dma_sync_single_for_cpu(&cp->dev, cp->rx_dma,
/* Now it is safe to examine the buffer. */
hp = (struct my_card_header *) cp->rx_buf;
if (header_is_ok(hp)) {
dma_unmap_single(&cp->dev, cp->rx_dma, cp->rx_len,
} else {
/* CPU should not write to
* DMA_FROM_DEVICE-mapped area,
* so dma_sync_single_for_device() is
* not needed here. It would be required
* for DMA_BIDIRECTIONAL mapping if
* the memory was modified.
Drivers converted fully to this interface should not use virt_to_bus any
longer, nor should they use bus_to_virt. Some drivers have to be changed a
little bit, because there is no longer an equivalent to bus_to_virt in the
dynamic DMA mapping scheme - you have to always store the DMA addresses
returned by the dma_alloc_coherent, dma_pool_alloc, and dma_map_single
calls (dma_map_sg stores them in the scatterlist itself if the platform
supports dynamic DMA mapping in hardware) in your driver structures and/or
in the card registers.
All drivers should be using these interfaces with no exceptions. It
is planned to completely remove virt_to_bus() and bus_to_virt() as
they are entirely deprecated. Some ports already do not provide these
as it is impossible to correctly support them.
Handling Errors
DMA address space is limited on some architectures and an allocation
failure can be determined by:
- checking if dma_alloc_coherent returns NULL or dma_map_sg returns 0
- checking the returned dma_addr_t of dma_map_single and dma_map_page
by using dma_mapping_error():
dma_addr_t dma_handle;
dma_handle = dma_map_single(dev, addr, size, direction);
if (dma_mapping_error(dev, dma_handle)) {
* reduce current DMA mapping usage,
* delay and try again later or
* reset driver.
Networking drivers must call dev_kfree_skb to free the socket buffer
and return NETDEV_TX_OK if the DMA mapping fails on the transmit hook
(ndo_start_xmit). This means that the socket buffer is just dropped in
the failure case.
SCSI drivers must return SCSI_MLQUEUE_HOST_BUSY if the DMA mapping
fails in the queuecommand hook. This means that the SCSI subsystem
passes the command to the driver again later.
Optimizing Unmap State Space Consumption
On many platforms, dma_unmap_{single,page}() is simply a nop.
Therefore, keeping track of the mapping address and length is a waste
of space. Instead of filling your drivers up with ifdefs and the like
to "work around" this (which would defeat the whole purpose of a
portable API) the following facilities are provided.
Actually, instead of describing the macros one by one, we'll
transform some example code.
1) Use DEFINE_DMA_UNMAP_{ADDR,LEN} in state saving structures.
Example, before:
struct ring_state {
struct sk_buff *skb;
dma_addr_t mapping;
__u32 len;
struct ring_state {
struct sk_buff *skb;
2) Use dma_unmap_{addr,len}_set to set these values.
Example, before:
ringp->mapping = FOO;
ringp->len = BAR;
dma_unmap_addr_set(ringp, mapping, FOO);
dma_unmap_len_set(ringp, len, BAR);
3) Use dma_unmap_{addr,len} to access these values.
Example, before:
dma_unmap_single(dev, ringp->mapping, ringp->len,
dma_unmap_addr(ringp, mapping),
dma_unmap_len(ringp, len),
It really should be self-explanatory. We treat the ADDR and LEN
separately, because it is possible for an implementation to only
need the address in order to perform the unmap operation.
Platform Issues
If you are just writing drivers for Linux and do not maintain
an architecture port for the kernel, you can safely skip down
to "Closing".
1) Struct scatterlist requirements.
Don't invent the architecture specific struct scatterlist; just use
<asm-generic/scatterlist.h>. You need to enable
CONFIG_NEED_SG_DMA_LENGTH if the architecture supports IOMMUs
(including software IOMMU).
Architectures must ensure that kmalloc'ed buffer is
DMA-safe. Drivers and subsystems depend on it. If an architecture
isn't fully DMA-coherent (i.e. hardware doesn't ensure that data in
the CPU cache is identical to data in main memory),
ARCH_DMA_MINALIGN must be set so that the memory allocator
makes sure that kmalloc'ed buffer doesn't share a cache line with
the others. See arch/arm/include/asm/cache.h as an example.
Note that ARCH_DMA_MINALIGN is about DMA memory alignment
constraints. You don't need to worry about the architecture data
alignment constraints (e.g. the alignment constraints about 64-bit
3) Supporting multiple types of IOMMUs
If your architecture needs to support multiple types of IOMMUs, you
can use include/linux/asm-generic/dma-mapping-common.h. It's a
library to support the DMA API with multiple types of IOMMUs. Lots
of architectures (x86, powerpc, sh, alpha, ia64, microblaze and
sparc) use it. Choose one to see how it can be used. If you need to
support multiple types of IOMMUs in a single system, the example of
x86 or powerpc helps.
This document, and the API itself, would not be in its current
form without the feedback and suggestions from numerous individuals.
We would like to specifically mention, in no particular order, the
following people:
Russell King <>
Leo Dagum <>
Ralf Baechle <>
Grant Grundler <>
Jay Estabrook <>
Thomas Sailer <>
Andrea Arcangeli <>
Jens Axboe <>
David Mosberger-Tang <>