DMA-API-HOWTO.txt 34.1 KB
Newer Older
1 2 3
=========================
Dynamic DMA mapping Guide
=========================
Linus Torvalds's avatar
Linus Torvalds committed
4

5 6 7
:Author: David S. Miller <davem@redhat.com>
:Author: Richard Henderson <rth@cygnus.com>
:Author: Jakub Jelinek <jakub@redhat.com>
Linus Torvalds's avatar
Linus Torvalds committed
8

9 10
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
Linus Torvalds's avatar
Linus Torvalds committed
11 12
DMA-API.txt.

13 14
CPU and DMA addresses
=====================
15 16 17 18 19 20

There are several kinds of addresses involved in the DMA API, and it's
important to understand the differences.

The kernel normally uses virtual addresses.  Any address returned by
kmalloc(), vmalloc(), and similar interfaces is a virtual address and can
21
be stored in a ``void *``.
22 23 24 25 26 27 28 29

The virtual memory system (TLB, page tables, etc.) translates virtual
addresses to CPU physical addresses, which are stored as "phys_addr_t" or
"resource_size_t".  The kernel manages device resources like registers as
physical addresses.  These are the addresses in /proc/iomem.  The physical
address is not directly useful to a driver; it must use ioremap() to map
the space and produce a virtual address.

Yinghai Lu's avatar
Yinghai Lu committed
30 31 32 33 34
I/O devices use a third kind of address: a "bus address".  If a device has
registers at an MMIO address, or if it performs DMA to read or write system
memory, the addresses used by the device are bus addresses.  In some
systems, bus addresses are identical to CPU physical addresses, but in
general they are not.  IOMMUs and host bridges can produce arbitrary
35 36
mappings between physical and bus addresses.

Yinghai Lu's avatar
Yinghai Lu committed
37 38 39 40 41
From a device's point of view, DMA uses the bus address space, but it may
be restricted to a subset of that space.  For example, even if a system
supports 64-bit addresses for main memory and PCI BARs, it may use an IOMMU
so devices only need to use 32-bit DMA addresses.

42
Here's a picture and some examples::
43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81

               CPU                  CPU                  Bus
             Virtual              Physical             Address
             Address              Address               Space
              Space                Space

            +-------+             +------+             +------+
            |       |             |MMIO  |   Offset    |      |
            |       |  Virtual    |Space |   applied   |      |
          C +-------+ --------> B +------+ ----------> +------+ A
            |       |  mapping    |      |   by host   |      |
  +-----+   |       |             |      |   bridge    |      |   +--------+
  |     |   |       |             +------+             |      |   |        |
  | CPU |   |       |             | RAM  |             |      |   | Device |
  |     |   |       |             |      |             |      |   |        |
  +-----+   +-------+             +------+             +------+   +--------+
            |       |  Virtual    |Buffer|   Mapping   |      |
          X +-------+ --------> Y +------+ <---------- +------+ Z
            |       |  mapping    | RAM  |   by IOMMU
            |       |             |      |
            |       |             |      |
            +-------+             +------+

During the enumeration process, the kernel learns about I/O devices and
their MMIO space and the host bridges that connect them to the system.  For
example, if a PCI device has a BAR, the kernel reads the bus address (A)
from the BAR and converts it to a CPU physical address (B).  The address B
is stored in a struct resource and usually exposed via /proc/iomem.  When a
driver claims a device, it typically uses ioremap() to map physical address
B at a virtual address (C).  It can then use, e.g., ioread32(C), to access
the device registers at bus address A.

If the device supports DMA, the driver sets up a buffer using kmalloc() or
a similar interface, which returns a virtual address (X).  The virtual
memory system maps X to a physical address (Y) in system RAM.  The driver
can use virtual address X to access the buffer, but the device itself
cannot because DMA doesn't go through the CPU virtual memory system.

In some simple systems, the device can do DMA directly to physical address
Yinghai Lu's avatar
Yinghai Lu committed
82
Y.  But in many others, there is IOMMU hardware that translates DMA
83 84 85
addresses to physical addresses, e.g., it translates Z to Y.  This is part
of the reason for the DMA API: the driver can give a virtual address X to
an interface like dma_map_single(), which sets up any required IOMMU
Yinghai Lu's avatar
Yinghai Lu committed
86
mapping and returns the DMA address Z.  The driver then tells the device to
87 88
do DMA to Z, and the IOMMU maps it to the buffer at address Y in system
RAM.
Linus Torvalds's avatar
Linus Torvalds committed
89 90 91 92 93 94 95

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
transfer.

The following API will work of course even on platforms where no such
96 97 98
hardware exists.

Note that the DMA API works with any bus independent of the underlying
99 100 101
microprocessor architecture. You should use the DMA API rather than the
bus-specific DMA API, i.e., use the dma_map_*() interfaces rather than the
pci_map_*() interfaces.
Linus Torvalds's avatar
Linus Torvalds committed
102

103
First of all, you should make sure::
Linus Torvalds's avatar
Linus Torvalds committed
104

105
	#include <linux/dma-mapping.h>
Linus Torvalds's avatar
Linus Torvalds committed
106

107
is in your driver, which provides the definition of dma_addr_t.  This type
Yinghai Lu's avatar
Yinghai Lu committed
108
can hold any valid DMA address for the platform and should be used
109
everywhere you hold a DMA address returned from the DMA mapping functions.
Linus Torvalds's avatar
Linus Torvalds committed
110

111 112
What memory is DMA'able?
========================
Linus Torvalds's avatar
Linus Torvalds committed
113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131

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. ]

David Brownell's avatar
David Brownell committed
132 133 134 135 136 137 138 139 140
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.)
Linus Torvalds's avatar
Linus Torvalds committed
141 142 143 144 145 146 147 148

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.

149 150
DMA addressing limitations
==========================
Linus Torvalds's avatar
Linus Torvalds committed
151 152

Does your device have any DMA addressing limitations?  For example, is
153 154
your device only capable of driving the low order 24-bits of address?
If so, you need to inform the kernel of this fact.
Linus Torvalds's avatar
Linus Torvalds committed
155 156

By default, the kernel assumes that your device can address the full
157 158 159 160 161 162 163 164 165 166 167 168 169
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,
Linus Torvalds's avatar
Linus Torvalds committed
170 171 172
because this shows that you did think about these issues wrt. your
device.

173
The query is performed via a call to dma_set_mask_and_coherent()::
Linus Torvalds's avatar
Linus Torvalds committed
174

175
	int dma_set_mask_and_coherent(struct device *dev, u64 mask);
Linus Torvalds's avatar
Linus Torvalds committed
176

177 178 179
which will query the mask for both streaming and coherent APIs together.
If you have some special requirements, then the following two separate
queries can be used instead:
Linus Torvalds's avatar
Linus Torvalds committed
180

181
	The query for streaming mappings is performed via a call to
182
	dma_set_mask()::
183 184 185 186

		int dma_set_mask(struct device *dev, u64 mask);

	The query for consistent allocations is performed via a call
187
	to dma_set_coherent_mask()::
188 189

		int dma_set_coherent_mask(struct device *dev, u64 mask);
Linus Torvalds's avatar
Linus Torvalds committed
190

191 192 193 194
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
195 196 197
device struct of your device is embedded in the bus-specific device
struct of your device.  For example, &pdev->dev is a pointer to the
device struct of a PCI device (pdev is a pointer to the PCI device
198
struct of your device).
Linus Torvalds's avatar
Linus Torvalds committed
199

200
If it returns non-zero, your device cannot perform DMA properly on
Linus Torvalds's avatar
Linus Torvalds committed
201 202 203 204 205 206 207 208 209 210 211 212 213 214 215
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.

216
The standard 32-bit addressing device would do something like this::
Linus Torvalds's avatar
Linus Torvalds committed
217

218
	if (dma_set_mask_and_coherent(dev, DMA_BIT_MASK(32))) {
219
		dev_warn(dev, "mydev: No suitable DMA available\n");
Linus Torvalds's avatar
Linus Torvalds committed
220 221 222
		goto ignore_this_device;
	}

223 224 225 226 227 228 229
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.
Linus Torvalds's avatar
Linus Torvalds committed
230 231

Here is how you would handle a 64-bit capable device which can drive
232
all 64-bits when accessing streaming DMA::
Linus Torvalds's avatar
Linus Torvalds committed
233 234 235

	int using_dac;

236
	if (!dma_set_mask(dev, DMA_BIT_MASK(64))) {
Linus Torvalds's avatar
Linus Torvalds committed
237
		using_dac = 1;
238
	} else if (!dma_set_mask(dev, DMA_BIT_MASK(32))) {
Linus Torvalds's avatar
Linus Torvalds committed
239 240
		using_dac = 0;
	} else {
241
		dev_warn(dev, "mydev: No suitable DMA available\n");
Linus Torvalds's avatar
Linus Torvalds committed
242 243 244 245
		goto ignore_this_device;
	}

If a card is capable of using 64-bit consistent allocations as well,
246
the case would look like this::
Linus Torvalds's avatar
Linus Torvalds committed
247 248 249

	int using_dac, consistent_using_dac;

250
	if (!dma_set_mask_and_coherent(dev, DMA_BIT_MASK(64))) {
Linus Torvalds's avatar
Linus Torvalds committed
251
		using_dac = 1;
252
		consistent_using_dac = 1;
253
	} else if (!dma_set_mask_and_coherent(dev, DMA_BIT_MASK(32))) {
Linus Torvalds's avatar
Linus Torvalds committed
254 255 256
		using_dac = 0;
		consistent_using_dac = 0;
	} else {
257
		dev_warn(dev, "mydev: No suitable DMA available\n");
Linus Torvalds's avatar
Linus Torvalds committed
258 259 260
		goto ignore_this_device;
	}

261 262 263 264
The coherent mask will always be able to set the same or a smaller mask as
the streaming 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().
Linus Torvalds's avatar
Linus Torvalds committed
265 266

Finally, if your device can only drive the low 24-bits of
267
address you might do something like::
Linus Torvalds's avatar
Linus Torvalds committed
268

269
	if (dma_set_mask(dev, DMA_BIT_MASK(24))) {
270
		dev_warn(dev, "mydev: 24-bit DMA addressing not available\n");
Linus Torvalds's avatar
Linus Torvalds committed
271 272 273
		goto ignore_this_device;
	}

274 275 276
When dma_set_mask() or dma_set_mask_and_coherent() 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.
Linus Torvalds's avatar
Linus Torvalds committed
277 278 279 280 281 282 283

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
284
is important that the last call to dma_set_mask() be for the
Linus Torvalds's avatar
Linus Torvalds committed
285 286
most specific mask.

287
Here is pseudo-code showing how this might be done::
Linus Torvalds's avatar
Linus Torvalds committed
288

289
	#define PLAYBACK_ADDRESS_BITS	DMA_BIT_MASK(32)
290
	#define RECORD_ADDRESS_BITS	DMA_BIT_MASK(24)
Linus Torvalds's avatar
Linus Torvalds committed
291 292

	struct my_sound_card *card;
293
	struct device *dev;
Linus Torvalds's avatar
Linus Torvalds committed
294 295

	...
296
	if (!dma_set_mask(dev, PLAYBACK_ADDRESS_BITS)) {
Linus Torvalds's avatar
Linus Torvalds committed
297 298 299
		card->playback_enabled = 1;
	} else {
		card->playback_enabled = 0;
300
		dev_warn(dev, "%s: Playback disabled due to DMA limitations\n",
Linus Torvalds's avatar
Linus Torvalds committed
301 302
		       card->name);
	}
303
	if (!dma_set_mask(dev, RECORD_ADDRESS_BITS)) {
Linus Torvalds's avatar
Linus Torvalds committed
304 305 306
		card->record_enabled = 1;
	} else {
		card->record_enabled = 0;
307
		dev_warn(dev, "%s: Record disabled due to DMA limitations\n",
Linus Torvalds's avatar
Linus Torvalds committed
308 309 310 311 312 313 314
		       card->name);
	}

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.

315 316
Types of DMA mappings
=====================
Linus Torvalds's avatar
Linus Torvalds committed
317 318 319 320 321 322 323 324 325 326 327 328

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
Yinghai Lu's avatar
Yinghai Lu committed
329
  bits of the DMA space.  However, for future compatibility you should
330
  set the consistent mask even if this default is fine for your
Linus Torvalds's avatar
Linus Torvalds committed
331 332 333 334 335 336 337 338 339 340 341 342 343
  driver.

  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.

344 345 346 347
  .. important::

	     Consistent DMA memory does not preclude the usage of
	     proper memory barriers.  The CPU may reorder stores to
Linus Torvalds's avatar
Linus Torvalds committed
348 349 350
	     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
351
	     something like::
Linus Torvalds's avatar
Linus Torvalds committed
352 353 354 355 356 357 358

		desc->word0 = address;
		wmb();
		desc->word1 = DESC_VALID;

             in order to get correct behavior on all platforms.

David Brownell's avatar
David Brownell committed
359 360 361 362 363
	     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).

364 365 366
- 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.
Linus Torvalds's avatar
Linus Torvalds committed
367

368
  Think of "streaming" as "asynchronous" or "outside the coherency
Linus Torvalds's avatar
Linus Torvalds committed
369 370 371 372 373 374 375 376 377 378 379 380
  domain".

  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.

381 382
Neither type of DMA mapping has alignment restrictions that come from
the underlying bus, although some devices may have such restrictions.
David Brownell's avatar
David Brownell committed
383 384 385
Also, systems with caches that aren't DMA-coherent will work better
when the underlying buffers don't share cache lines with other data.

Linus Torvalds's avatar
Linus Torvalds committed
386

387 388
Using Consistent DMA mappings
=============================
Linus Torvalds's avatar
Linus Torvalds committed
389 390

To allocate and map large (PAGE_SIZE or so) consistent DMA regions,
391
you should do::
Linus Torvalds's avatar
Linus Torvalds committed
392 393 394

	dma_addr_t dma_handle;

395
	cpu_addr = dma_alloc_coherent(dev, size, &dma_handle, gfp);
Linus Torvalds's avatar
Linus Torvalds committed
396

397
where device is a ``struct device *``. This may be called in interrupt
398
context with the GFP_ATOMIC flag.
Linus Torvalds's avatar
Linus Torvalds committed
399 400 401 402

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
403
__get_free_pages() (but takes size instead of a page order).  If your
Linus Torvalds's avatar
Linus Torvalds committed
404
driver needs regions sized smaller than a page, you may prefer using
405 406 407 408 409 410 411 412 413 414
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
well.

415
dma_alloc_coherent() returns two values: the virtual address which you
Linus Torvalds's avatar
Linus Torvalds committed
416 417 418
can use to access it from the CPU and dma_handle which you pass to the
card.

Yinghai Lu's avatar
Yinghai Lu committed
419
The CPU virtual address and the DMA address are both
Linus Torvalds's avatar
Linus Torvalds committed
420 421 422 423 424 425
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.

426
To unmap and free such a DMA region, you call::
Linus Torvalds's avatar
Linus Torvalds committed
427

428
	dma_free_coherent(dev, size, cpu_addr, dma_handle);
Linus Torvalds's avatar
Linus Torvalds committed
429

430
where dev, size are the same as in the above call and cpu_addr and
431
dma_handle are the values dma_alloc_coherent() returned to you.
Linus Torvalds's avatar
Linus Torvalds committed
432 433 434
This function may not be called in interrupt context.

If your driver needs lots of smaller memory regions, you can write
435
custom code to subdivide pages returned by dma_alloc_coherent(),
436
or you can use the dma_pool API to do that.  A dma_pool is like
437
a kmem_cache, but it uses dma_alloc_coherent(), not __get_free_pages().
Linus Torvalds's avatar
Linus Torvalds committed
438 439 440
Also, it understands common hardware constraints for alignment,
like queue heads needing to be aligned on N byte boundaries.

441
Create a dma_pool like this::
Linus Torvalds's avatar
Linus Torvalds committed
442

443
	struct dma_pool *pool;
Linus Torvalds's avatar
Linus Torvalds committed
444

445
	pool = dma_pool_create(name, dev, size, align, boundary);
Linus Torvalds's avatar
Linus Torvalds committed
446

447
The "name" is for diagnostics (like a kmem_cache name); dev and size
Linus Torvalds's avatar
Linus Torvalds committed
448 449 450
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,
451
pass 0 for boundary; passing 4096 says memory allocated from this pool
Linus Torvalds's avatar
Linus Torvalds committed
452
must not cross 4KByte boundaries (but at that time it may be better to
453
use dma_alloc_coherent() directly instead).
Linus Torvalds's avatar
Linus Torvalds committed
454

455
Allocate memory from a DMA pool like this::
Linus Torvalds's avatar
Linus Torvalds committed
456

457
	cpu_addr = dma_pool_alloc(pool, flags, &dma_handle);
Linus Torvalds's avatar
Linus Torvalds committed
458

459 460
flags are GFP_KERNEL if blocking is permitted (not in_interrupt nor
holding SMP locks), GFP_ATOMIC otherwise.  Like dma_alloc_coherent(),
Linus Torvalds's avatar
Linus Torvalds committed
461 462
this returns two values, cpu_addr and dma_handle.

463
Free memory that was allocated from a dma_pool like this::
Linus Torvalds's avatar
Linus Torvalds committed
464

465
	dma_pool_free(pool, cpu_addr, dma_handle);
Linus Torvalds's avatar
Linus Torvalds committed
466

467 468
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
Linus Torvalds's avatar
Linus Torvalds committed
469 470
may be called in interrupt context.

471
Destroy a dma_pool by calling::
Linus Torvalds's avatar
Linus Torvalds committed
472

473
	dma_pool_destroy(pool);
Linus Torvalds's avatar
Linus Torvalds committed
474

475
Make sure you've called dma_pool_free() for all memory allocated
Linus Torvalds's avatar
Linus Torvalds committed
476 477 478
from a pool before you destroy the pool. This function may not
be called in interrupt context.

479 480
DMA Direction
=============
Linus Torvalds's avatar
Linus Torvalds committed
481 482 483

The interfaces described in subsequent portions of this document
take a DMA direction argument, which is an integer and takes on
484
one of the following values::
Linus Torvalds's avatar
Linus Torvalds committed
485

486 487 488 489
 DMA_BIDIRECTIONAL
 DMA_TO_DEVICE
 DMA_FROM_DEVICE
 DMA_NONE
Linus Torvalds's avatar
Linus Torvalds committed
490

491
You should provide the exact DMA direction if you know it.
Linus Torvalds's avatar
Linus Torvalds committed
492

493 494
DMA_TO_DEVICE means "from main memory to the device"
DMA_FROM_DEVICE means "from the device to main memory"
Linus Torvalds's avatar
Linus Torvalds committed
495 496 497 498 499 500 501
It is the direction in which the data moves during the DMA
transfer.

You are _strongly_ encouraged to specify this as precisely
as you possibly can.

If you absolutely cannot know the direction of the DMA transfer,
502
specify DMA_BIDIRECTIONAL.  It means that the DMA can go in
Linus Torvalds's avatar
Linus Torvalds committed
503 504 505 506
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.

507
The value DMA_NONE is to be used for debugging.  One can
Linus Torvalds's avatar
Linus Torvalds committed
508 509 510 511 512 513 514 515 516
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
517
kernel logs when the DMA controller hardware detects violation of the
Linus Torvalds's avatar
Linus Torvalds committed
518 519 520 521
permission setting.

Only streaming mappings specify a direction, consistent mappings
implicitly have a direction attribute setting of
522
DMA_BIDIRECTIONAL.
Linus Torvalds's avatar
Linus Torvalds committed
523

524 525 526
The SCSI subsystem tells you the direction to use in the
'sc_data_direction' member of the SCSI command your driver is
working on.
Linus Torvalds's avatar
Linus Torvalds committed
527 528

For Networking drivers, it's a rather simple affair.  For transmit
529
packets, map/unmap them with the DMA_TO_DEVICE direction
Linus Torvalds's avatar
Linus Torvalds committed
530
specifier.  For receive packets, just the opposite, map/unmap them
531
with the DMA_FROM_DEVICE direction specifier.
Linus Torvalds's avatar
Linus Torvalds committed
532

533 534
Using Streaming DMA mappings
============================
Linus Torvalds's avatar
Linus Torvalds committed
535 536 537 538 539 540

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
scatterlist.

541
To map a single region, you do::
Linus Torvalds's avatar
Linus Torvalds committed
542

543
	struct device *dev = &my_dev->dev;
Linus Torvalds's avatar
Linus Torvalds committed
544 545 546 547
	dma_addr_t dma_handle;
	void *addr = buffer->ptr;
	size_t size = buffer->len;

548
	dma_handle = dma_map_single(dev, addr, size, direction);
549
	if (dma_mapping_error(dev, dma_handle)) {
550 551 552 553 554 555 556
		/*
		 * reduce current DMA mapping usage,
		 * delay and try again later or
		 * reset driver.
		 */
		goto map_error_handling;
	}
Linus Torvalds's avatar
Linus Torvalds committed
557

558
and to unmap it::
Linus Torvalds's avatar
Linus Torvalds committed
559

560
	dma_unmap_single(dev, dma_handle, size, direction);
Linus Torvalds's avatar
Linus Torvalds committed
561

562
You should call dma_mapping_error() as dma_map_single() could fail and return
563 564 565 566 567
error.  Doing so will ensure that the mapping code will work correctly on all
DMA implementations without any dependency on the specifics of the underlying
implementation. Using the returned address without checking for errors could
result in failures ranging from panics to silent data corruption.  The same
applies to dma_map_page() as well.
568

569
You should call dma_unmap_single() when the DMA activity is finished, e.g.,
Linus Torvalds's avatar
Linus Torvalds committed
570 571
from the interrupt which told you that the DMA transfer is done.

572
Using CPU pointers like this for single mappings has a disadvantage:
Linus Torvalds's avatar
Linus Torvalds committed
573
you cannot reference HIGHMEM memory in this way.  Thus, there is a
574
map/unmap interface pair akin to dma_{map,unmap}_single().  These
575
interfaces deal with page/offset pairs instead of CPU pointers.
576
Specifically::
Linus Torvalds's avatar
Linus Torvalds committed
577

578
	struct device *dev = &my_dev->dev;
Linus Torvalds's avatar
Linus Torvalds committed
579 580 581 582 583
	dma_addr_t dma_handle;
	struct page *page = buffer->page;
	unsigned long offset = buffer->offset;
	size_t size = buffer->len;

584
	dma_handle = dma_map_page(dev, page, offset, size, direction);
585
	if (dma_mapping_error(dev, dma_handle)) {
586 587 588 589 590 591 592
		/*
		 * reduce current DMA mapping usage,
		 * delay and try again later or
		 * reset driver.
		 */
		goto map_error_handling;
	}
Linus Torvalds's avatar
Linus Torvalds committed
593 594 595

	...

596
	dma_unmap_page(dev, dma_handle, size, direction);
Linus Torvalds's avatar
Linus Torvalds committed
597 598 599

Here, "offset" means byte offset within the given page.

600 601 602
You should call dma_mapping_error() as dma_map_page() could fail and return
error as outlined under the dma_map_single() discussion.

603
You should call dma_unmap_page() when the DMA activity is finished, e.g.,
604 605
from the interrupt which told you that the DMA transfer is done.

606
With scatterlists, you map a region gathered from several regions by::
Linus Torvalds's avatar
Linus Torvalds committed
607

608
	int i, count = dma_map_sg(dev, sglist, nents, direction);
Linus Torvalds's avatar
Linus Torvalds committed
609 610
	struct scatterlist *sg;

611
	for_each_sg(sglist, sg, count, i) {
Linus Torvalds's avatar
Linus Torvalds committed
612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629
		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.

630
To unmap a scatterlist, just call::
Linus Torvalds's avatar
Linus Torvalds committed
631

632
	dma_unmap_sg(dev, sglist, nents, direction);
Linus Torvalds's avatar
Linus Torvalds committed
633 634 635

Again, make sure DMA activity has already finished.

636 637 638 639 640 641
.. 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.
Linus Torvalds's avatar
Linus Torvalds committed
642

643
Every dma_map_{single,sg}() call should have its dma_unmap_{single,sg}()
Yinghai Lu's avatar
Yinghai Lu committed
644 645
counterpart, because the DMA address space is a shared resource and
you could render the machine unusable by consuming all DMA addresses.
Linus Torvalds's avatar
Linus Torvalds committed
646 647 648

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
649
properly in order for the CPU and device to see the most up-to-date and
Linus Torvalds's avatar
Linus Torvalds committed
650 651
correct copy of the DMA buffer.

652
So, firstly, just map it with dma_map_{single,sg}(), and after each DMA
653
transfer call either::
Linus Torvalds's avatar
Linus Torvalds committed
654

655
	dma_sync_single_for_cpu(dev, dma_handle, size, direction);
Linus Torvalds's avatar
Linus Torvalds committed
656

657
or::
Linus Torvalds's avatar
Linus Torvalds committed
658

659
	dma_sync_sg_for_cpu(dev, sglist, nents, direction);
Linus Torvalds's avatar
Linus Torvalds committed
660 661 662 663

as appropriate.

Then, if you wish to let the device get at the DMA area again,
664
finish accessing the data with the CPU, and then before actually
665
giving the buffer to the hardware call either::
Linus Torvalds's avatar
Linus Torvalds committed
666

667
	dma_sync_single_for_device(dev, dma_handle, size, direction);
Linus Torvalds's avatar
Linus Torvalds committed
668

669
or::
Linus Torvalds's avatar
Linus Torvalds committed
670

671
	dma_sync_sg_for_device(dev, sglist, nents, direction);
Linus Torvalds's avatar
Linus Torvalds committed
672 673 674

as appropriate.

675 676 677
.. note::

	      The 'nents' argument to dma_sync_sg_for_cpu() and
678 679 680 681
	      dma_sync_sg_for_device() must be the same passed to
	      dma_map_sg(). It is _NOT_ the count returned by
	      dma_map_sg().

Linus Torvalds's avatar
Linus Torvalds committed
682
After the last DMA transfer call one of the DMA unmap routines
683 684 685
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.
Linus Torvalds's avatar
Linus Torvalds committed
686 687

Here is pseudo code which shows a situation in which you would need
688
to use the dma_sync_*() interfaces::
Linus Torvalds's avatar
Linus Torvalds committed
689 690 691 692 693

	my_card_setup_receive_buffer(struct my_card *cp, char *buffer, int len)
	{
		dma_addr_t mapping;

694
		mapping = dma_map_single(cp->dev, buffer, len, DMA_FROM_DEVICE);
695
		if (dma_mapping_error(cp->dev, mapping)) {
696 697 698 699 700 701 702
			/*
			 * reduce current DMA mapping usage,
			 * delay and try again later or
			 * reset driver.
			 */
			goto map_error_handling;
		}
Linus Torvalds's avatar
Linus Torvalds committed
703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725

		cp->rx_buf = buffer;
		cp->rx_len = len;
		cp->rx_dma = mapping;

		give_rx_buf_to_card(cp);
	}

	...

	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.
			 */
726 727 728
			dma_sync_single_for_cpu(&cp->dev, cp->rx_dma,
						cp->rx_len,
						DMA_FROM_DEVICE);
Linus Torvalds's avatar
Linus Torvalds committed
729 730 731 732

			/* Now it is safe to examine the buffer. */
			hp = (struct my_card_header *) cp->rx_buf;
			if (header_is_ok(hp)) {
733 734
				dma_unmap_single(&cp->dev, cp->rx_dma, cp->rx_len,
						 DMA_FROM_DEVICE);
Linus Torvalds's avatar
Linus Torvalds committed
735 736 737
				pass_to_upper_layers(cp->rx_buf);
				make_and_setup_new_rx_buf(cp);
			} else {
738 739 740 741 742 743
				/* 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.
Linus Torvalds's avatar
Linus Torvalds committed
744 745 746 747 748 749
				 */
				give_rx_buf_to_card(cp);
			}
		}
	}

750 751 752
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
Linus Torvalds's avatar
Linus Torvalds committed
753
dynamic DMA mapping scheme - you have to always store the DMA addresses
754 755
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
Linus Torvalds's avatar
Linus Torvalds committed
756 757 758
supports dynamic DMA mapping in hardware) in your driver structures and/or
in the card registers.

759 760
All drivers should be using these interfaces with no exceptions.  It
is planned to completely remove virt_to_bus() and bus_to_virt() as
Linus Torvalds's avatar
Linus Torvalds committed
761 762 763
they are entirely deprecated.  Some ports already do not provide these
as it is impossible to correctly support them.

764 765
Handling Errors
===============
766 767 768 769

DMA address space is limited on some architectures and an allocation
failure can be determined by:

770
- checking if dma_alloc_coherent() returns NULL or dma_map_sg returns 0
771

772
- checking the dma_addr_t returned from dma_map_single() and dma_map_page()
773
  by using dma_mapping_error()::
774 775 776 777 778 779 780 781 782 783

	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.
		 */
784 785 786 787 788 789 790
		goto map_error_handling;
	}

- unmap pages that are already mapped, when mapping error occurs in the middle
  of a multiple page mapping attempt. These example are applicable to
  dma_map_page() as well.

791 792
Example 1::

793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820
	dma_addr_t dma_handle1;
	dma_addr_t dma_handle2;

	dma_handle1 = dma_map_single(dev, addr, size, direction);
	if (dma_mapping_error(dev, dma_handle1)) {
		/*
		 * reduce current DMA mapping usage,
		 * delay and try again later or
		 * reset driver.
		 */
		goto map_error_handling1;
	}
	dma_handle2 = dma_map_single(dev, addr, size, direction);
	if (dma_mapping_error(dev, dma_handle2)) {
		/*
		 * reduce current DMA mapping usage,
		 * delay and try again later or
		 * reset driver.
		 */
		goto map_error_handling2;
	}

	...

	map_error_handling2:
		dma_unmap_single(dma_handle1);
	map_error_handling1:

821 822 823 824 825 826
Example 2::

	/*
	 * if buffers are allocated in a loop, unmap all mapped buffers when
	 * mapping error is detected in the middle
	 */
827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857

	dma_addr_t dma_addr;
	dma_addr_t array[DMA_BUFFERS];
	int save_index = 0;

	for (i = 0; i < DMA_BUFFERS; i++) {

		...

		dma_addr = dma_map_single(dev, addr, size, direction);
		if (dma_mapping_error(dev, dma_addr)) {
			/*
			 * reduce current DMA mapping usage,
			 * delay and try again later or
			 * reset driver.
			 */
			goto map_error_handling;
		}
		array[i].dma_addr = dma_addr;
		save_index++;
	}

	...

	map_error_handling:

	for (i = 0; i < save_index; i++) {

		...

		dma_unmap_single(array[i].dma_addr);
858 859
	}

860
Networking drivers must call dev_kfree_skb() to free the socket buffer
861 862 863 864 865 866 867 868
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.

869 870
Optimizing Unmap State Space Consumption
========================================
Linus Torvalds's avatar
Linus Torvalds committed
871

872
On many platforms, dma_unmap_{single,page}() is simply a nop.
Linus Torvalds's avatar
Linus Torvalds committed
873 874 875 876 877 878 879 880
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.

881
1) Use DEFINE_DMA_UNMAP_{ADDR,LEN} in state saving structures.
882
   Example, before::
Linus Torvalds's avatar
Linus Torvalds committed
883 884 885 886 887 888 889

	struct ring_state {
		struct sk_buff *skb;
		dma_addr_t mapping;
		__u32 len;
	};

890
   after::
Linus Torvalds's avatar
Linus Torvalds committed
891 892 893

	struct ring_state {
		struct sk_buff *skb;
894 895
		DEFINE_DMA_UNMAP_ADDR(mapping);
		DEFINE_DMA_UNMAP_LEN(len);
Linus Torvalds's avatar
Linus Torvalds committed
896 897
	};

898
2) Use dma_unmap_{addr,len}_set() to set these values.
899
   Example, before::
Linus Torvalds's avatar
Linus Torvalds committed
900 901 902 903

	ringp->mapping = FOO;
	ringp->len = BAR;

904
   after::
Linus Torvalds's avatar
Linus Torvalds committed
905

906 907
	dma_unmap_addr_set(ringp, mapping, FOO);
	dma_unmap_len_set(ringp, len, BAR);
Linus Torvalds's avatar
Linus Torvalds committed
908

909
3) Use dma_unmap_{addr,len}() to access these values.
910
   Example, before::
Linus Torvalds's avatar
Linus Torvalds committed
911

912 913
	dma_unmap_single(dev, ringp->mapping, ringp->len,
			 DMA_FROM_DEVICE);
Linus Torvalds's avatar
Linus Torvalds committed
914

915
   after::
Linus Torvalds's avatar
Linus Torvalds committed
916

917 918 919 920
	dma_unmap_single(dev,
			 dma_unmap_addr(ringp, mapping),
			 dma_unmap_len(ringp, len),
			 DMA_FROM_DEVICE);
Linus Torvalds's avatar
Linus Torvalds committed
921 922 923 924 925

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.

926 927
Platform Issues
===============
Linus Torvalds's avatar
Linus Torvalds committed
928 929 930 931 932 933 934

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.

935 936
   You need to enable CONFIG_NEED_SG_DMA_LENGTH if the architecture
   supports IOMMUs (including software IOMMU).
Linus Torvalds's avatar
Linus Torvalds committed
937

938
2) ARCH_DMA_MINALIGN
939 940 941 942 943

   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),
944
   ARCH_DMA_MINALIGN must be set so that the memory allocator
945 946 947
   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.

948
   Note that ARCH_DMA_MINALIGN is about DMA memory alignment
949 950 951
   constraints. You don't need to worry about the architecture data
   alignment constraints (e.g. the alignment constraints about 64-bit
   objects).
Linus Torvalds's avatar
Linus Torvalds committed
952

953 954
Closing
=======
Linus Torvalds's avatar
Linus Torvalds committed
955

956
This document, and the API itself, would not be in its current
Linus Torvalds's avatar
Linus Torvalds committed
957 958
form without the feedback and suggestions from numerous individuals.
We would like to specifically mention, in no particular order, the
959
following people::
Linus Torvalds's avatar
Linus Torvalds committed
960 961 962 963 964 965 966 967

	Russell King <rmk@arm.linux.org.uk>
	Leo Dagum <dagum@barrel.engr.sgi.com>
	Ralf Baechle <ralf@oss.sgi.com>
	Grant Grundler <grundler@cup.hp.com>
	Jay Estabrook <Jay.Estabrook@compaq.com>
	Thomas Sailer <sailer@ife.ee.ethz.ch>
	Andrea Arcangeli <andrea@suse.de>
968
	Jens Axboe <jens.axboe@oracle.com>
Linus Torvalds's avatar
Linus Torvalds committed
969
	David Mosberger-Tang <davidm@hpl.hp.com>