aboutsummaryrefslogtreecommitdiffstats
path: root/Documentation/DMA-mapping.txt
diff options
context:
space:
mode:
authorLinus Torvalds <torvalds@ppc970.osdl.org>2005-04-16 18:20:36 -0400
committerLinus Torvalds <torvalds@ppc970.osdl.org>2005-04-16 18:20:36 -0400
commit1da177e4c3f41524e886b7f1b8a0c1fc7321cac2 (patch)
tree0bba044c4ce775e45a88a51686b5d9f90697ea9d /Documentation/DMA-mapping.txt
Linux-2.6.12-rc2v2.6.12-rc2
Initial git repository build. I'm not bothering with the full history, even though we have it. We can create a separate "historical" git archive of that later if we want to, and in the meantime it's about 3.2GB when imported into git - space that would just make the early git days unnecessarily complicated, when we don't have a lot of good infrastructure for it. Let it rip!
Diffstat (limited to 'Documentation/DMA-mapping.txt')
-rw-r--r--Documentation/DMA-mapping.txt881
1 files changed, 881 insertions, 0 deletions
diff --git a/Documentation/DMA-mapping.txt b/Documentation/DMA-mapping.txt
new file mode 100644
index 000000000000..f4ac37f157ea
--- /dev/null
+++ b/Documentation/DMA-mapping.txt
@@ -0,0 +1,881 @@
1 Dynamic DMA mapping
2 ===================
3
4 David S. Miller <davem@redhat.com>
5 Richard Henderson <rth@cygnus.com>
6 Jakub Jelinek <jakub@redhat.com>
7
8This document describes the DMA mapping system in terms of the pci_
9API. For a similar API that works for generic devices, see
10DMA-API.txt.
11
12Most of the 64bit platforms have special hardware that translates bus
13addresses (DMA addresses) into physical addresses. This is similar to
14how page tables and/or a TLB translates virtual addresses to physical
15addresses on a CPU. This is needed so that e.g. PCI devices can
16access with a Single Address Cycle (32bit DMA address) any page in the
1764bit physical address space. Previously in Linux those 64bit
18platforms had to set artificial limits on the maximum RAM size in the
19system, so that the virt_to_bus() static scheme works (the DMA address
20translation tables were simply filled on bootup to map each bus
21address to the physical page __pa(bus_to_virt())).
22
23So that Linux can use the dynamic DMA mapping, it needs some help from the
24drivers, namely it has to take into account that DMA addresses should be
25mapped only for the time they are actually used and unmapped after the DMA
26transfer.
27
28The following API will work of course even on platforms where no such
29hardware exists, see e.g. include/asm-i386/pci.h for how it is implemented on
30top of the virt_to_bus interface.
31
32First of all, you should make sure
33
34#include <linux/pci.h>
35
36is in your driver. This file will obtain for you the definition of the
37dma_addr_t (which can hold any valid DMA address for the platform)
38type which should be used everywhere you hold a DMA (bus) address
39returned from the DMA mapping functions.
40
41 What memory is DMA'able?
42
43The first piece of information you must know is what kernel memory can
44be used with the DMA mapping facilities. There has been an unwritten
45set of rules regarding this, and this text is an attempt to finally
46write them down.
47
48If you acquired your memory via the page allocator
49(i.e. __get_free_page*()) or the generic memory allocators
50(i.e. kmalloc() or kmem_cache_alloc()) then you may DMA to/from
51that memory using the addresses returned from those routines.
52
53This means specifically that you may _not_ use the memory/addresses
54returned from vmalloc() for DMA. It is possible to DMA to the
55_underlying_ memory mapped into a vmalloc() area, but this requires
56walking page tables to get the physical addresses, and then
57translating each of those pages back to a kernel address using
58something like __va(). [ EDIT: Update this when we integrate
59Gerd Knorr's generic code which does this. ]
60
61This rule also means that you may not use kernel image addresses
62(ie. items in the kernel's data/text/bss segment, or your driver's)
63nor may you use kernel stack addresses for DMA. Both of these items
64might be mapped somewhere entirely different than the rest of physical
65memory.
66
67Also, this means that you cannot take the return of a kmap()
68call and DMA to/from that. This is similar to vmalloc().
69
70What about block I/O and networking buffers? The block I/O and
71networking subsystems make sure that the buffers they use are valid
72for you to DMA from/to.
73
74 DMA addressing limitations
75
76Does your device have any DMA addressing limitations? For example, is
77your device only capable of driving the low order 24-bits of address
78on the PCI bus for SAC DMA transfers? If so, you need to inform the
79PCI layer of this fact.
80
81By default, the kernel assumes that your device can address the full
8232-bits in a SAC cycle. For a 64-bit DAC capable device, this needs
83to be increased. And for a device with limitations, as discussed in
84the previous paragraph, it needs to be decreased.
85
86pci_alloc_consistent() by default will return 32-bit DMA addresses.
87PCI-X specification requires PCI-X devices to support 64-bit
88addressing (DAC) for all transactions. And at least one platform (SGI
89SN2) requires 64-bit consistent allocations to operate correctly when
90the IO bus is in PCI-X mode. Therefore, like with pci_set_dma_mask(),
91it's good practice to call pci_set_consistent_dma_mask() to set the
92appropriate mask even if your device only supports 32-bit DMA
93(default) and especially if it's a PCI-X device.
94
95For correct operation, you must interrogate the PCI layer in your
96device probe routine to see if the PCI controller on the machine can
97properly support the DMA addressing limitation your device has. It is
98good style to do this even if your device holds the default setting,
99because this shows that you did think about these issues wrt. your
100device.
101
102The query is performed via a call to pci_set_dma_mask():
103
104 int pci_set_dma_mask(struct pci_dev *pdev, u64 device_mask);
105
106The query for consistent allocations is performed via a a call to
107pci_set_consistent_dma_mask():
108
109 int pci_set_consistent_dma_mask(struct pci_dev *pdev, u64 device_mask);
110
111Here, pdev is a pointer to the PCI device struct of your device, and
112device_mask is a bit mask describing which bits of a PCI address your
113device supports. It returns zero if your card can perform DMA
114properly on the machine given the address mask you provided.
115
116If it returns non-zero, your device can not perform DMA properly on
117this platform, and attempting to do so will result in undefined
118behavior. You must either use a different mask, or not use DMA.
119
120This means that in the failure case, you have three options:
121
1221) Use another DMA mask, if possible (see below).
1232) Use some non-DMA mode for data transfer, if possible.
1243) Ignore this device and do not initialize it.
125
126It is recommended that your driver print a kernel KERN_WARNING message
127when you end up performing either #2 or #3. In this manner, if a user
128of your driver reports that performance is bad or that the device is not
129even detected, you can ask them for the kernel messages to find out
130exactly why.
131
132The standard 32-bit addressing PCI device would do something like
133this:
134
135 if (pci_set_dma_mask(pdev, DMA_32BIT_MASK)) {
136 printk(KERN_WARNING
137 "mydev: No suitable DMA available.\n");
138 goto ignore_this_device;
139 }
140
141Another common scenario is a 64-bit capable device. The approach
142here is to try for 64-bit DAC addressing, but back down to a
14332-bit mask should that fail. The PCI platform code may fail the
14464-bit mask not because the platform is not capable of 64-bit
145addressing. Rather, it may fail in this case simply because
14632-bit SAC addressing is done more efficiently than DAC addressing.
147Sparc64 is one platform which behaves in this way.
148
149Here is how you would handle a 64-bit capable device which can drive
150all 64-bits when accessing streaming DMA:
151
152 int using_dac;
153
154 if (!pci_set_dma_mask(pdev, DMA_64BIT_MASK)) {
155 using_dac = 1;
156 } else if (!pci_set_dma_mask(pdev, DMA_32BIT_MASK)) {
157 using_dac = 0;
158 } else {
159 printk(KERN_WARNING
160 "mydev: No suitable DMA available.\n");
161 goto ignore_this_device;
162 }
163
164If a card is capable of using 64-bit consistent allocations as well,
165the case would look like this:
166
167 int using_dac, consistent_using_dac;
168
169 if (!pci_set_dma_mask(pdev, DMA_64BIT_MASK)) {
170 using_dac = 1;
171 consistent_using_dac = 1;
172 pci_set_consistent_dma_mask(pdev, DMA_64BIT_MASK);
173 } else if (!pci_set_dma_mask(pdev, DMA_32BIT_MASK)) {
174 using_dac = 0;
175 consistent_using_dac = 0;
176 pci_set_consistent_dma_mask(pdev, DMA_32BIT_MASK);
177 } else {
178 printk(KERN_WARNING
179 "mydev: No suitable DMA available.\n");
180 goto ignore_this_device;
181 }
182
183pci_set_consistent_dma_mask() will always be able to set the same or a
184smaller mask as pci_set_dma_mask(). However for the rare case that a
185device driver only uses consistent allocations, one would have to
186check the return value from pci_set_consistent_dma_mask().
187
188If your 64-bit device is going to be an enormous consumer of DMA
189mappings, this can be problematic since the DMA mappings are a
190finite resource on many platforms. Please see the "DAC Addressing
191for Address Space Hungry Devices" section near the end of this
192document for how to handle this case.
193
194Finally, if your device can only drive the low 24-bits of
195address during PCI bus mastering you might do something like:
196
197 if (pci_set_dma_mask(pdev, 0x00ffffff)) {
198 printk(KERN_WARNING
199 "mydev: 24-bit DMA addressing not available.\n");
200 goto ignore_this_device;
201 }
202
203When pci_set_dma_mask() is successful, and returns zero, the PCI layer
204saves away this mask you have provided. The PCI layer will use this
205information later when you make DMA mappings.
206
207There is a case which we are aware of at this time, which is worth
208mentioning in this documentation. If your device supports multiple
209functions (for example a sound card provides playback and record
210functions) and the various different functions have _different_
211DMA addressing limitations, you may wish to probe each mask and
212only provide the functionality which the machine can handle. It
213is important that the last call to pci_set_dma_mask() be for the
214most specific mask.
215
216Here is pseudo-code showing how this might be done:
217
218 #define PLAYBACK_ADDRESS_BITS DMA_32BIT_MASK
219 #define RECORD_ADDRESS_BITS 0x00ffffff
220
221 struct my_sound_card *card;
222 struct pci_dev *pdev;
223
224 ...
225 if (!pci_set_dma_mask(pdev, PLAYBACK_ADDRESS_BITS)) {
226 card->playback_enabled = 1;
227 } else {
228 card->playback_enabled = 0;
229 printk(KERN_WARN "%s: Playback disabled due to DMA limitations.\n",
230 card->name);
231 }
232 if (!pci_set_dma_mask(pdev, RECORD_ADDRESS_BITS)) {
233 card->record_enabled = 1;
234 } else {
235 card->record_enabled = 0;
236 printk(KERN_WARN "%s: Record disabled due to DMA limitations.\n",
237 card->name);
238 }
239
240A sound card was used as an example here because this genre of PCI
241devices seems to be littered with ISA chips given a PCI front end,
242and thus retaining the 16MB DMA addressing limitations of ISA.
243
244 Types of DMA mappings
245
246There are two types of DMA mappings:
247
248- Consistent DMA mappings which are usually mapped at driver
249 initialization, unmapped at the end and for which the hardware should
250 guarantee that the device and the CPU can access the data
251 in parallel and will see updates made by each other without any
252 explicit software flushing.
253
254 Think of "consistent" as "synchronous" or "coherent".
255
256 The current default is to return consistent memory in the low 32
257 bits of the PCI bus space. However, for future compatibility you
258 should set the consistent mask even if this default is fine for your
259 driver.
260
261 Good examples of what to use consistent mappings for are:
262
263 - Network card DMA ring descriptors.
264 - SCSI adapter mailbox command data structures.
265 - Device firmware microcode executed out of
266 main memory.
267
268 The invariant these examples all require is that any CPU store
269 to memory is immediately visible to the device, and vice
270 versa. Consistent mappings guarantee this.
271
272 IMPORTANT: Consistent DMA memory does not preclude the usage of
273 proper memory barriers. The CPU may reorder stores to
274 consistent memory just as it may normal memory. Example:
275 if it is important for the device to see the first word
276 of a descriptor updated before the second, you must do
277 something like:
278
279 desc->word0 = address;
280 wmb();
281 desc->word1 = DESC_VALID;
282
283 in order to get correct behavior on all platforms.
284
285- Streaming DMA mappings which are usually mapped for one DMA transfer,
286 unmapped right after it (unless you use pci_dma_sync_* below) and for which
287 hardware can optimize for sequential accesses.
288
289 This of "streaming" as "asynchronous" or "outside the coherency
290 domain".
291
292 Good examples of what to use streaming mappings for are:
293
294 - Networking buffers transmitted/received by a device.
295 - Filesystem buffers written/read by a SCSI device.
296
297 The interfaces for using this type of mapping were designed in
298 such a way that an implementation can make whatever performance
299 optimizations the hardware allows. To this end, when using
300 such mappings you must be explicit about what you want to happen.
301
302Neither type of DMA mapping has alignment restrictions that come
303from PCI, although some devices may have such restrictions.
304
305 Using Consistent DMA mappings.
306
307To allocate and map large (PAGE_SIZE or so) consistent DMA regions,
308you should do:
309
310 dma_addr_t dma_handle;
311
312 cpu_addr = pci_alloc_consistent(dev, size, &dma_handle);
313
314where dev is a struct pci_dev *. You should pass NULL for PCI like buses
315where devices don't have struct pci_dev (like ISA, EISA). This may be
316called in interrupt context.
317
318This argument is needed because the DMA translations may be bus
319specific (and often is private to the bus which the device is attached
320to).
321
322Size is the length of the region you want to allocate, in bytes.
323
324This routine will allocate RAM for that region, so it acts similarly to
325__get_free_pages (but takes size instead of a page order). If your
326driver needs regions sized smaller than a page, you may prefer using
327the pci_pool interface, described below.
328
329The consistent DMA mapping interfaces, for non-NULL dev, will by
330default return a DMA address which is SAC (Single Address Cycle)
331addressable. Even if the device indicates (via PCI dma mask) that it
332may address the upper 32-bits and thus perform DAC cycles, consistent
333allocation will only return > 32-bit PCI addresses for DMA if the
334consistent dma mask has been explicitly changed via
335pci_set_consistent_dma_mask(). This is true of the pci_pool interface
336as well.
337
338pci_alloc_consistent returns two values: the virtual address which you
339can use to access it from the CPU and dma_handle which you pass to the
340card.
341
342The cpu return address and the DMA bus master address are both
343guaranteed to be aligned to the smallest PAGE_SIZE order which
344is greater than or equal to the requested size. This invariant
345exists (for example) to guarantee that if you allocate a chunk
346which is smaller than or equal to 64 kilobytes, the extent of the
347buffer you receive will not cross a 64K boundary.
348
349To unmap and free such a DMA region, you call:
350
351 pci_free_consistent(dev, size, cpu_addr, dma_handle);
352
353where dev, size are the same as in the above call and cpu_addr and
354dma_handle are the values pci_alloc_consistent returned to you.
355This function may not be called in interrupt context.
356
357If your driver needs lots of smaller memory regions, you can write
358custom code to subdivide pages returned by pci_alloc_consistent,
359or you can use the pci_pool API to do that. A pci_pool is like
360a kmem_cache, but it uses pci_alloc_consistent not __get_free_pages.
361Also, it understands common hardware constraints for alignment,
362like queue heads needing to be aligned on N byte boundaries.
363
364Create a pci_pool like this:
365
366 struct pci_pool *pool;
367
368 pool = pci_pool_create(name, dev, size, align, alloc);
369
370The "name" is for diagnostics (like a kmem_cache name); dev and size
371are as above. The device's hardware alignment requirement for this
372type of data is "align" (which is expressed in bytes, and must be a
373power of two). If your device has no boundary crossing restrictions,
374pass 0 for alloc; passing 4096 says memory allocated from this pool
375must not cross 4KByte boundaries (but at that time it may be better to
376go for pci_alloc_consistent directly instead).
377
378Allocate memory from a pci pool like this:
379
380 cpu_addr = pci_pool_alloc(pool, flags, &dma_handle);
381
382flags are SLAB_KERNEL if blocking is permitted (not in_interrupt nor
383holding SMP locks), SLAB_ATOMIC otherwise. Like pci_alloc_consistent,
384this returns two values, cpu_addr and dma_handle.
385
386Free memory that was allocated from a pci_pool like this:
387
388 pci_pool_free(pool, cpu_addr, dma_handle);
389
390where pool is what you passed to pci_pool_alloc, and cpu_addr and
391dma_handle are the values pci_pool_alloc returned. This function
392may be called in interrupt context.
393
394Destroy a pci_pool by calling:
395
396 pci_pool_destroy(pool);
397
398Make sure you've called pci_pool_free for all memory allocated
399from a pool before you destroy the pool. This function may not
400be called in interrupt context.
401
402 DMA Direction
403
404The interfaces described in subsequent portions of this document
405take a DMA direction argument, which is an integer and takes on
406one of the following values:
407
408 PCI_DMA_BIDIRECTIONAL
409 PCI_DMA_TODEVICE
410 PCI_DMA_FROMDEVICE
411 PCI_DMA_NONE
412
413One should provide the exact DMA direction if you know it.
414
415PCI_DMA_TODEVICE means "from main memory to the PCI device"
416PCI_DMA_FROMDEVICE means "from the PCI device to main memory"
417It is the direction in which the data moves during the DMA
418transfer.
419
420You are _strongly_ encouraged to specify this as precisely
421as you possibly can.
422
423If you absolutely cannot know the direction of the DMA transfer,
424specify PCI_DMA_BIDIRECTIONAL. It means that the DMA can go in
425either direction. The platform guarantees that you may legally
426specify this, and that it will work, but this may be at the
427cost of performance for example.
428
429The value PCI_DMA_NONE is to be used for debugging. One can
430hold this in a data structure before you come to know the
431precise direction, and this will help catch cases where your
432direction tracking logic has failed to set things up properly.
433
434Another advantage of specifying this value precisely (outside of
435potential platform-specific optimizations of such) is for debugging.
436Some platforms actually have a write permission boolean which DMA
437mappings can be marked with, much like page protections in the user
438program address space. Such platforms can and do report errors in the
439kernel logs when the PCI controller hardware detects violation of the
440permission setting.
441
442Only streaming mappings specify a direction, consistent mappings
443implicitly have a direction attribute setting of
444PCI_DMA_BIDIRECTIONAL.
445
446The SCSI subsystem provides mechanisms for you to easily obtain
447the direction to use, in the SCSI command:
448
449 scsi_to_pci_dma_dir(SCSI_DIRECTION)
450
451Where SCSI_DIRECTION is obtained from the 'sc_data_direction'
452member of the SCSI command your driver is working on. The
453mentioned interface above returns a value suitable for passing
454into the streaming DMA mapping interfaces below.
455
456For Networking drivers, it's a rather simple affair. For transmit
457packets, map/unmap them with the PCI_DMA_TODEVICE direction
458specifier. For receive packets, just the opposite, map/unmap them
459with the PCI_DMA_FROMDEVICE direction specifier.
460
461 Using Streaming DMA mappings
462
463The streaming DMA mapping routines can be called from interrupt
464context. There are two versions of each map/unmap, one which will
465map/unmap a single memory region, and one which will map/unmap a
466scatterlist.
467
468To map a single region, you do:
469
470 struct pci_dev *pdev = mydev->pdev;
471 dma_addr_t dma_handle;
472 void *addr = buffer->ptr;
473 size_t size = buffer->len;
474
475 dma_handle = pci_map_single(dev, addr, size, direction);
476
477and to unmap it:
478
479 pci_unmap_single(dev, dma_handle, size, direction);
480
481You should call pci_unmap_single when the DMA activity is finished, e.g.
482from the interrupt which told you that the DMA transfer is done.
483
484Using cpu pointers like this for single mappings has a disadvantage,
485you cannot reference HIGHMEM memory in this way. Thus, there is a
486map/unmap interface pair akin to pci_{map,unmap}_single. These
487interfaces deal with page/offset pairs instead of cpu pointers.
488Specifically:
489
490 struct pci_dev *pdev = mydev->pdev;
491 dma_addr_t dma_handle;
492 struct page *page = buffer->page;
493 unsigned long offset = buffer->offset;
494 size_t size = buffer->len;
495
496 dma_handle = pci_map_page(dev, page, offset, size, direction);
497
498 ...
499
500 pci_unmap_page(dev, dma_handle, size, direction);
501
502Here, "offset" means byte offset within the given page.
503
504With scatterlists, you map a region gathered from several regions by:
505
506 int i, count = pci_map_sg(dev, sglist, nents, direction);
507 struct scatterlist *sg;
508
509 for (i = 0, sg = sglist; i < count; i++, sg++) {
510 hw_address[i] = sg_dma_address(sg);
511 hw_len[i] = sg_dma_len(sg);
512 }
513
514where nents is the number of entries in the sglist.
515
516The implementation is free to merge several consecutive sglist entries
517into one (e.g. if DMA mapping is done with PAGE_SIZE granularity, any
518consecutive sglist entries can be merged into one provided the first one
519ends and the second one starts on a page boundary - in fact this is a huge
520advantage for cards which either cannot do scatter-gather or have very
521limited number of scatter-gather entries) and returns the actual number
522of sg entries it mapped them to. On failure 0 is returned.
523
524Then you should loop count times (note: this can be less than nents times)
525and use sg_dma_address() and sg_dma_len() macros where you previously
526accessed sg->address and sg->length as shown above.
527
528To unmap a scatterlist, just call:
529
530 pci_unmap_sg(dev, sglist, nents, direction);
531
532Again, make sure DMA activity has already finished.
533
534PLEASE NOTE: The 'nents' argument to the pci_unmap_sg call must be
535 the _same_ one you passed into the pci_map_sg call,
536 it should _NOT_ be the 'count' value _returned_ from the
537 pci_map_sg call.
538
539Every pci_map_{single,sg} call should have its pci_unmap_{single,sg}
540counterpart, because the bus address space is a shared resource (although
541in some ports the mapping is per each BUS so less devices contend for the
542same bus address space) and you could render the machine unusable by eating
543all bus addresses.
544
545If you need to use the same streaming DMA region multiple times and touch
546the data in between the DMA transfers, the buffer needs to be synced
547properly in order for the cpu and device to see the most uptodate and
548correct copy of the DMA buffer.
549
550So, firstly, just map it with pci_map_{single,sg}, and after each DMA
551transfer call either:
552
553 pci_dma_sync_single_for_cpu(dev, dma_handle, size, direction);
554
555or:
556
557 pci_dma_sync_sg_for_cpu(dev, sglist, nents, direction);
558
559as appropriate.
560
561Then, if you wish to let the device get at the DMA area again,
562finish accessing the data with the cpu, and then before actually
563giving the buffer to the hardware call either:
564
565 pci_dma_sync_single_for_device(dev, dma_handle, size, direction);
566
567or:
568
569 pci_dma_sync_sg_for_device(dev, sglist, nents, direction);
570
571as appropriate.
572
573After the last DMA transfer call one of the DMA unmap routines
574pci_unmap_{single,sg}. If you don't touch the data from the first pci_map_*
575call till pci_unmap_*, then you don't have to call the pci_dma_sync_*
576routines at all.
577
578Here is pseudo code which shows a situation in which you would need
579to use the pci_dma_sync_*() interfaces.
580
581 my_card_setup_receive_buffer(struct my_card *cp, char *buffer, int len)
582 {
583 dma_addr_t mapping;
584
585 mapping = pci_map_single(cp->pdev, buffer, len, PCI_DMA_FROMDEVICE);
586
587 cp->rx_buf = buffer;
588 cp->rx_len = len;
589 cp->rx_dma = mapping;
590
591 give_rx_buf_to_card(cp);
592 }
593
594 ...
595
596 my_card_interrupt_handler(int irq, void *devid, struct pt_regs *regs)
597 {
598 struct my_card *cp = devid;
599
600 ...
601 if (read_card_status(cp) == RX_BUF_TRANSFERRED) {
602 struct my_card_header *hp;
603
604 /* Examine the header to see if we wish
605 * to accept the data. But synchronize
606 * the DMA transfer with the CPU first
607 * so that we see updated contents.
608 */
609 pci_dma_sync_single_for_cpu(cp->pdev, cp->rx_dma,
610 cp->rx_len,
611 PCI_DMA_FROMDEVICE);
612
613 /* Now it is safe to examine the buffer. */
614 hp = (struct my_card_header *) cp->rx_buf;
615 if (header_is_ok(hp)) {
616 pci_unmap_single(cp->pdev, cp->rx_dma, cp->rx_len,
617 PCI_DMA_FROMDEVICE);
618 pass_to_upper_layers(cp->rx_buf);
619 make_and_setup_new_rx_buf(cp);
620 } else {
621 /* Just sync the buffer and give it back
622 * to the card.
623 */
624 pci_dma_sync_single_for_device(cp->pdev,
625 cp->rx_dma,
626 cp->rx_len,
627 PCI_DMA_FROMDEVICE);
628 give_rx_buf_to_card(cp);
629 }
630 }
631 }
632
633Drivers converted fully to this interface should not use virt_to_bus any
634longer, nor should they use bus_to_virt. Some drivers have to be changed a
635little bit, because there is no longer an equivalent to bus_to_virt in the
636dynamic DMA mapping scheme - you have to always store the DMA addresses
637returned by the pci_alloc_consistent, pci_pool_alloc, and pci_map_single
638calls (pci_map_sg stores them in the scatterlist itself if the platform
639supports dynamic DMA mapping in hardware) in your driver structures and/or
640in the card registers.
641
642All PCI drivers should be using these interfaces with no exceptions.
643It is planned to completely remove virt_to_bus() and bus_to_virt() as
644they are entirely deprecated. Some ports already do not provide these
645as it is impossible to correctly support them.
646
647 64-bit DMA and DAC cycle support
648
649Do you understand all of the text above? Great, then you already
650know how to use 64-bit DMA addressing under Linux. Simply make
651the appropriate pci_set_dma_mask() calls based upon your cards
652capabilities, then use the mapping APIs above.
653
654It is that simple.
655
656Well, not for some odd devices. See the next section for information
657about that.
658
659 DAC Addressing for Address Space Hungry Devices
660
661There exists a class of devices which do not mesh well with the PCI
662DMA mapping API. By definition these "mappings" are a finite
663resource. The number of total available mappings per bus is platform
664specific, but there will always be a reasonable amount.
665
666What is "reasonable"? Reasonable means that networking and block I/O
667devices need not worry about using too many mappings.
668
669As an example of a problematic device, consider compute cluster cards.
670They can potentially need to access gigabytes of memory at once via
671DMA. Dynamic mappings are unsuitable for this kind of access pattern.
672
673To this end we've provided a small API by which a device driver
674may use DAC cycles to directly address all of physical memory.
675Not all platforms support this, but most do. It is easy to determine
676whether the platform will work properly at probe time.
677
678First, understand that there may be a SEVERE performance penalty for
679using these interfaces on some platforms. Therefore, you MUST only
680use these interfaces if it is absolutely required. %99 of devices can
681use the normal APIs without any problems.
682
683Note that for streaming type mappings you must either use these
684interfaces, or the dynamic mapping interfaces above. You may not mix
685usage of both for the same device. Such an act is illegal and is
686guaranteed to put a banana in your tailpipe.
687
688However, consistent mappings may in fact be used in conjunction with
689these interfaces. Remember that, as defined, consistent mappings are
690always going to be SAC addressable.
691
692The first thing your driver needs to do is query the PCI platform
693layer with your devices DAC addressing capabilities:
694
695 int pci_dac_set_dma_mask(struct pci_dev *pdev, u64 mask);
696
697This routine behaves identically to pci_set_dma_mask. You may not
698use the following interfaces if this routine fails.
699
700Next, DMA addresses using this API are kept track of using the
701dma64_addr_t type. It is guaranteed to be big enough to hold any
702DAC address the platform layer will give to you from the following
703routines. If you have consistent mappings as well, you still
704use plain dma_addr_t to keep track of those.
705
706All mappings obtained here will be direct. The mappings are not
707translated, and this is the purpose of this dialect of the DMA API.
708
709All routines work with page/offset pairs. This is the _ONLY_ way to
710portably refer to any piece of memory. If you have a cpu pointer
711(which may be validly DMA'd too) you may easily obtain the page
712and offset using something like this:
713
714 struct page *page = virt_to_page(ptr);
715 unsigned long offset = offset_in_page(ptr);
716
717Here are the interfaces:
718
719 dma64_addr_t pci_dac_page_to_dma(struct pci_dev *pdev,
720 struct page *page,
721 unsigned long offset,
722 int direction);
723
724The DAC address for the tuple PAGE/OFFSET are returned. The direction
725argument is the same as for pci_{map,unmap}_single(). The same rules
726for cpu/device access apply here as for the streaming mapping
727interfaces. To reiterate:
728
729 The cpu may touch the buffer before pci_dac_page_to_dma.
730 The device may touch the buffer after pci_dac_page_to_dma
731 is made, but the cpu may NOT.
732
733When the DMA transfer is complete, invoke:
734
735 void pci_dac_dma_sync_single_for_cpu(struct pci_dev *pdev,
736 dma64_addr_t dma_addr,
737 size_t len, int direction);
738
739This must be done before the CPU looks at the buffer again.
740This interface behaves identically to pci_dma_sync_{single,sg}_for_cpu().
741
742And likewise, if you wish to let the device get back at the buffer after
743the cpu has read/written it, invoke:
744
745 void pci_dac_dma_sync_single_for_device(struct pci_dev *pdev,
746 dma64_addr_t dma_addr,
747 size_t len, int direction);
748
749before letting the device access the DMA area again.
750
751If you need to get back to the PAGE/OFFSET tuple from a dma64_addr_t
752the following interfaces are provided:
753
754 struct page *pci_dac_dma_to_page(struct pci_dev *pdev,
755 dma64_addr_t dma_addr);
756 unsigned long pci_dac_dma_to_offset(struct pci_dev *pdev,
757 dma64_addr_t dma_addr);
758
759This is possible with the DAC interfaces purely because they are
760not translated in any way.
761
762 Optimizing Unmap State Space Consumption
763
764On many platforms, pci_unmap_{single,page}() is simply a nop.
765Therefore, keeping track of the mapping address and length is a waste
766of space. Instead of filling your drivers up with ifdefs and the like
767to "work around" this (which would defeat the whole purpose of a
768portable API) the following facilities are provided.
769
770Actually, instead of describing the macros one by one, we'll
771transform some example code.
772
7731) Use DECLARE_PCI_UNMAP_{ADDR,LEN} in state saving structures.
774 Example, before:
775
776 struct ring_state {
777 struct sk_buff *skb;
778 dma_addr_t mapping;
779 __u32 len;
780 };
781
782 after:
783
784 struct ring_state {
785 struct sk_buff *skb;
786 DECLARE_PCI_UNMAP_ADDR(mapping)
787 DECLARE_PCI_UNMAP_LEN(len)
788 };
789
790 NOTE: DO NOT put a semicolon at the end of the DECLARE_*()
791 macro.
792
7932) Use pci_unmap_{addr,len}_set to set these values.
794 Example, before:
795
796 ringp->mapping = FOO;
797 ringp->len = BAR;
798
799 after:
800
801 pci_unmap_addr_set(ringp, mapping, FOO);
802 pci_unmap_len_set(ringp, len, BAR);
803
8043) Use pci_unmap_{addr,len} to access these values.
805 Example, before:
806
807 pci_unmap_single(pdev, ringp->mapping, ringp->len,
808 PCI_DMA_FROMDEVICE);
809
810 after:
811
812 pci_unmap_single(pdev,
813 pci_unmap_addr(ringp, mapping),
814 pci_unmap_len(ringp, len),
815 PCI_DMA_FROMDEVICE);
816
817It really should be self-explanatory. We treat the ADDR and LEN
818separately, because it is possible for an implementation to only
819need the address in order to perform the unmap operation.
820
821 Platform Issues
822
823If you are just writing drivers for Linux and do not maintain
824an architecture port for the kernel, you can safely skip down
825to "Closing".
826
8271) Struct scatterlist requirements.
828
829 Struct scatterlist must contain, at a minimum, the following
830 members:
831
832 struct page *page;
833 unsigned int offset;
834 unsigned int length;
835
836 The base address is specified by a "page+offset" pair.
837
838 Previous versions of struct scatterlist contained a "void *address"
839 field that was sometimes used instead of page+offset. As of Linux
840 2.5., page+offset is always used, and the "address" field has been
841 deleted.
842
8432) More to come...
844
845 Handling Errors
846
847DMA address space is limited on some architectures and an allocation
848failure can be determined by:
849
850- checking if pci_alloc_consistent returns NULL or pci_map_sg returns 0
851
852- checking the returned dma_addr_t of pci_map_single and pci_map_page
853 by using pci_dma_mapping_error():
854
855 dma_addr_t dma_handle;
856
857 dma_handle = pci_map_single(dev, addr, size, direction);
858 if (pci_dma_mapping_error(dma_handle)) {
859 /*
860 * reduce current DMA mapping usage,
861 * delay and try again later or
862 * reset driver.
863 */
864 }
865
866 Closing
867
868This document, and the API itself, would not be in it's current
869form without the feedback and suggestions from numerous individuals.
870We would like to specifically mention, in no particular order, the
871following people:
872
873 Russell King <rmk@arm.linux.org.uk>
874 Leo Dagum <dagum@barrel.engr.sgi.com>
875 Ralf Baechle <ralf@oss.sgi.com>
876 Grant Grundler <grundler@cup.hp.com>
877 Jay Estabrook <Jay.Estabrook@compaq.com>
878 Thomas Sailer <sailer@ife.ee.ethz.ch>
879 Andrea Arcangeli <andrea@suse.de>
880 Jens Axboe <axboe@suse.de>
881 David Mosberger-Tang <davidm@hpl.hp.com>