1               Dynamic DMA mapping using the generic device
   2               ============================================
   4        James E.J. Bottomley <>
   6This document describes the DMA API.  For a more gentle introduction
   7of the API (and actual examples) see
  10This API is split into two pieces.  Part I describes the API.  Part II
  11describes the extensions to the API for supporting non-consistent
  12memory machines.  Unless you know that your driver absolutely has to
  13support non-consistent platforms (this is usually only legacy
  14platforms) you should only use the API described in part I.
  16Part I - dma_ API
  19To get the dma_ API, you must #include <linux/dma-mapping.h>
  22Part Ia - Using large dma-coherent buffers
  25void *
  26dma_alloc_coherent(struct device *dev, size_t size,
  27                             dma_addr_t *dma_handle, gfp_t flag)
  29Consistent memory is memory for which a write by either the device or
  30the processor can immediately be read by the processor or device
  31without having to worry about caching effects.  (You may however need
  32to make sure to flush the processor's write buffers before telling
  33devices to read that memory.)
  35This routine allocates a region of <size> bytes of consistent memory.
  36It also returns a <dma_handle> which may be cast to an unsigned
  37integer the same width as the bus and used as the physical address
  38base of the region.
  40Returns: a pointer to the allocated region (in the processor's virtual
  41address space) or NULL if the allocation failed.
  43Note: consistent memory can be expensive on some platforms, and the
  44minimum allocation length may be as big as a page, so you should
  45consolidate your requests for consistent memory as much as possible.
  46The simplest way to do that is to use the dma_pool calls (see below).
  48The flag parameter (dma_alloc_coherent only) allows the caller to
  49specify the GFP_ flags (see kmalloc) for the allocation (the
  50implementation may choose to ignore flags that affect the location of
  51the returned memory, like GFP_DMA).
  53void *
  54dma_zalloc_coherent(struct device *dev, size_t size,
  55                             dma_addr_t *dma_handle, gfp_t flag)
  57Wraps dma_alloc_coherent() and also zeroes the returned memory if the
  58allocation attempt succeeded.
  61dma_free_coherent(struct device *dev, size_t size, void *cpu_addr,
  62                           dma_addr_t dma_handle)
  64Free the region of consistent memory you previously allocated.  dev,
  65size and dma_handle must all be the same as those passed into the
  66consistent allocate.  cpu_addr must be the virtual address returned by
  67the consistent allocate.
  69Note that unlike their sibling allocation calls, these routines
  70may only be called with IRQs enabled.
  73Part Ib - Using small dma-coherent buffers
  76To get this part of the dma_ API, you must #include <linux/dmapool.h>
  78Many drivers need lots of small dma-coherent memory regions for DMA
  79descriptors or I/O buffers.  Rather than allocating in units of a page
  80or more using dma_alloc_coherent(), you can use DMA pools.  These work
  81much like a struct kmem_cache, except that they use the dma-coherent allocator,
  82not __get_free_pages().  Also, they understand common hardware constraints
  83for alignment, like queue heads needing to be aligned on N-byte boundaries.
  86        struct dma_pool *
  87        dma_pool_create(const char *name, struct device *dev,
  88                        size_t size, size_t align, size_t alloc);
  90The pool create() routines initialize a pool of dma-coherent buffers
  91for use with a given device.  It must be called in a context which
  92can sleep.
  94The "name" is for diagnostics (like a struct kmem_cache name); dev and size
  95are like what you'd pass to dma_alloc_coherent().  The device's hardware
  96alignment requirement for this type of data is "align" (which is expressed
  97in bytes, and must be a power of two).  If your device has no boundary
  98crossing restrictions, pass 0 for alloc; passing 4096 says memory allocated
  99from this pool must not cross 4KByte boundaries.
 102        void *dma_pool_alloc(struct dma_pool *pool, gfp_t gfp_flags,
 103                        dma_addr_t *dma_handle);
 105This allocates memory from the pool; the returned memory will meet the size
 106and alignment requirements specified at creation time.  Pass GFP_ATOMIC to
 107prevent blocking, or if it's permitted (not in_interrupt, not holding SMP locks),
 108pass GFP_KERNEL to allow blocking.  Like dma_alloc_coherent(), this returns
 109two values:  an address usable by the cpu, and the dma address usable by the
 110pool's device.
 113        void dma_pool_free(struct dma_pool *pool, void *vaddr,
 114                        dma_addr_t addr);
 116This puts memory back into the pool.  The pool is what was passed to
 117the pool allocation routine; the cpu (vaddr) and dma addresses are what
 118were returned when that routine allocated the memory being freed.
 121        void dma_pool_destroy(struct dma_pool *pool);
 123The pool destroy() routines free the resources of the pool.  They must be
 124called in a context which can sleep.  Make sure you've freed all allocated
 125memory back to the pool before you destroy it.
 128Part Ic - DMA addressing limitations
 132dma_supported(struct device *dev, u64 mask)
 134Checks to see if the device can support DMA to the memory described by
 137Returns: 1 if it can and 0 if it can't.
 139Notes: This routine merely tests to see if the mask is possible.  It
 140won't change the current mask settings.  It is more intended as an
 141internal API for use by the platform than an external API for use by
 142driver writers.
 145dma_set_mask(struct device *dev, u64 mask)
 147Checks to see if the mask is possible and updates the device
 148parameters if it is.
 150Returns: 0 if successful and a negative error if not.
 153dma_set_coherent_mask(struct device *dev, u64 mask)
 155Checks to see if the mask is possible and updates the device
 156parameters if it is.
 158Returns: 0 if successful and a negative error if not.
 161dma_get_required_mask(struct device *dev)
 163This API returns the mask that the platform requires to
 164operate efficiently.  Usually this means the returned mask
 165is the minimum required to cover all of memory.  Examining the
 166required mask gives drivers with variable descriptor sizes the
 167opportunity to use smaller descriptors as necessary.
 169Requesting the required mask does not alter the current mask.  If you
 170wish to take advantage of it, you should issue a dma_set_mask()
 171call to set the mask to the value returned.
 174Part Id - Streaming DMA mappings
 178dma_map_single(struct device *dev, void *cpu_addr, size_t size,
 179                      enum dma_data_direction direction)
 181Maps a piece of processor virtual memory so it can be accessed by the
 182device and returns the physical handle of the memory.
 184The direction for both api's may be converted freely by casting.
 185However the dma_ API uses a strongly typed enumerator for its
 188DMA_NONE                no direction (used for debugging)
 189DMA_TO_DEVICE           data is going from the memory to the device
 190DMA_FROM_DEVICE         data is coming from the device to the memory
 191DMA_BIDIRECTIONAL       direction isn't known
 193Notes:  Not all memory regions in a machine can be mapped by this
 194API.  Further, regions that appear to be physically contiguous in
 195kernel virtual space may not be contiguous as physical memory.  Since
 196this API does not provide any scatter/gather capability, it will fail
 197if the user tries to map a non-physically contiguous piece of memory.
 198For this reason, it is recommended that memory mapped by this API be
 199obtained only from sources which guarantee it to be physically contiguous
 200(like kmalloc).
 202Further, the physical address of the memory must be within the
 203dma_mask of the device (the dma_mask represents a bit mask of the
 204addressable region for the device.  I.e., if the physical address of
 205the memory anded with the dma_mask is still equal to the physical
 206address, then the device can perform DMA to the memory).  In order to
 207ensure that the memory allocated by kmalloc is within the dma_mask,
 208the driver may specify various platform-dependent flags to restrict
 209the physical memory range of the allocation (e.g. on x86, GFP_DMA
 210guarantees to be within the first 16Mb of available physical memory,
 211as required by ISA devices).
 213Note also that the above constraints on physical contiguity and
 214dma_mask may not apply if the platform has an IOMMU (a device which
 215supplies a physical to virtual mapping between the I/O memory bus and
 216the device).  However, to be portable, device driver writers may *not*
 217assume that such an IOMMU exists.
 219Warnings:  Memory coherency operates at a granularity called the cache
 220line width.  In order for memory mapped by this API to operate
 221correctly, the mapped region must begin exactly on a cache line
 222boundary and end exactly on one (to prevent two separately mapped
 223regions from sharing a single cache line).  Since the cache line size
 224may not be known at compile time, the API will not enforce this
 225requirement.  Therefore, it is recommended that driver writers who
 226don't take special care to determine the cache line size at run time
 227only map virtual regions that begin and end on page boundaries (which
 228are guaranteed also to be cache line boundaries).
 230DMA_TO_DEVICE synchronisation must be done after the last modification
 231of the memory region by the software and before it is handed off to
 232the driver.  Once this primitive is used, memory covered by this
 233primitive should be treated as read-only by the device.  If the device
 234may write to it at any point, it should be DMA_BIDIRECTIONAL (see
 237DMA_FROM_DEVICE synchronisation must be done before the driver
 238accesses data that may be changed by the device.  This memory should
 239be treated as read-only by the driver.  If the driver needs to write
 240to it at any point, it should be DMA_BIDIRECTIONAL (see below).
 242DMA_BIDIRECTIONAL requires special handling: it means that the driver
 243isn't sure if the memory was modified before being handed off to the
 244device and also isn't sure if the device will also modify it.  Thus,
 245you must always sync bidirectional memory twice: once before the
 246memory is handed off to the device (to make sure all memory changes
 247are flushed from the processor) and once before the data may be
 248accessed after being used by the device (to make sure any processor
 249cache lines are updated with data that the device may have changed).
 252dma_unmap_single(struct device *dev, dma_addr_t dma_addr, size_t size,
 253                 enum dma_data_direction direction)
 255Unmaps the region previously mapped.  All the parameters passed in
 256must be identical to those passed in (and returned) by the mapping
 260dma_map_page(struct device *dev, struct page *page,
 261                    unsigned long offset, size_t size,
 262                    enum dma_data_direction direction)
 264dma_unmap_page(struct device *dev, dma_addr_t dma_address, size_t size,
 265               enum dma_data_direction direction)
 267API for mapping and unmapping for pages.  All the notes and warnings
 268for the other mapping APIs apply here.  Also, although the <offset>
 269and <size> parameters are provided to do partial page mapping, it is
 270recommended that you never use these unless you really know what the
 271cache width is.
 274dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
 276In some circumstances dma_map_single and dma_map_page will fail to create
 277a mapping. A driver can check for these errors by testing the returned
 278dma address with dma_mapping_error(). A non-zero return value means the mapping
 279could not be created and the driver should take appropriate action (e.g.
 280reduce current DMA mapping usage or delay and try again later).
 282        int
 283        dma_map_sg(struct device *dev, struct scatterlist *sg,
 284                int nents, enum dma_data_direction direction)
 286Returns: the number of physical segments mapped (this may be shorter
 287than <nents> passed in if some elements of the scatter/gather list are
 288physically or virtually adjacent and an IOMMU maps them with a single
 291Please note that the sg cannot be mapped again if it has been mapped once.
 292The mapping process is allowed to destroy information in the sg.
 294As with the other mapping interfaces, dma_map_sg can fail. When it
 295does, 0 is returned and a driver must take appropriate action. It is
 296critical that the driver do something, in the case of a block driver
 297aborting the request or even oopsing is better than doing nothing and
 298corrupting the filesystem.
 300With scatterlists, you use the resulting mapping like this:
 302        int i, count = dma_map_sg(dev, sglist, nents, direction);
 303        struct scatterlist *sg;
 305        for_each_sg(sglist, sg, count, i) {
 306                hw_address[i] = sg_dma_address(sg);
 307                hw_len[i] = sg_dma_len(sg);
 308        }
 310where nents is the number of entries in the sglist.
 312The implementation is free to merge several consecutive sglist entries
 313into one (e.g. with an IOMMU, or if several pages just happen to be
 314physically contiguous) and returns the actual number of sg entries it
 315mapped them to. On failure 0, is returned.
 317Then you should loop count times (note: this can be less than nents times)
 318and use sg_dma_address() and sg_dma_len() macros where you previously
 319accessed sg->address and sg->length as shown above.
 321        void
 322        dma_unmap_sg(struct device *dev, struct scatterlist *sg,
 323                int nhwentries, enum dma_data_direction direction)
 325Unmap the previously mapped scatter/gather list.  All the parameters
 326must be the same as those and passed in to the scatter/gather mapping
 329Note: <nents> must be the number you passed in, *not* the number of
 330physical entries returned.
 333dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle, size_t size,
 334                        enum dma_data_direction direction)
 336dma_sync_single_for_device(struct device *dev, dma_addr_t dma_handle, size_t size,
 337                           enum dma_data_direction direction)
 339dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg, int nelems,
 340                    enum dma_data_direction direction)
 342dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg, int nelems,
 343                       enum dma_data_direction direction)
 345Synchronise a single contiguous or scatter/gather mapping for the cpu
 346and device. With the sync_sg API, all the parameters must be the same
 347as those passed into the single mapping API. With the sync_single API,
 348you can use dma_handle and size parameters that aren't identical to
 349those passed into the single mapping API to do a partial sync.
 351Notes:  You must do this:
 353- Before reading values that have been written by DMA from the device
 354  (use the DMA_FROM_DEVICE direction)
 355- After writing values that will be written to the device using DMA
 356  (use the DMA_TO_DEVICE) direction
 357- before *and* after handing memory to the device if the memory is
 360See also dma_map_single().
 363dma_map_single_attrs(struct device *dev, void *cpu_addr, size_t size,
 364                     enum dma_data_direction dir,
 365                     struct dma_attrs *attrs)
 368dma_unmap_single_attrs(struct device *dev, dma_addr_t dma_addr,
 369                       size_t size, enum dma_data_direction dir,
 370                       struct dma_attrs *attrs)
 373dma_map_sg_attrs(struct device *dev, struct scatterlist *sgl,
 374                 int nents, enum dma_data_direction dir,
 375                 struct dma_attrs *attrs)
 378dma_unmap_sg_attrs(struct device *dev, struct scatterlist *sgl,
 379                   int nents, enum dma_data_direction dir,
 380                   struct dma_attrs *attrs)
 382The four functions above are just like the counterpart functions
 383without the _attrs suffixes, except that they pass an optional
 384struct dma_attrs*.
 386struct dma_attrs encapsulates a set of "dma attributes". For the
 387definition of struct dma_attrs see linux/dma-attrs.h.
 389The interpretation of dma attributes is architecture-specific, and
 390each attribute should be documented in Documentation/DMA-attributes.txt.
 392If struct dma_attrs* is NULL, the semantics of each of these
 393functions is identical to those of the corresponding function
 394without the _attrs suffix. As a result dma_map_single_attrs()
 395can generally replace dma_map_single(), etc.
 397As an example of the use of the *_attrs functions, here's how
 398you could pass an attribute DMA_ATTR_FOO when mapping memory
 399for DMA:
 401#include <linux/dma-attrs.h>
 402/* DMA_ATTR_FOO should be defined in linux/dma-attrs.h and
 403 * documented in Documentation/DMA-attributes.txt */
 406        DEFINE_DMA_ATTRS(attrs);
 407        dma_set_attr(DMA_ATTR_FOO, &attrs);
 408        ....
 409        n = dma_map_sg_attrs(dev, sg, nents, DMA_TO_DEVICE, &attr);
 410        ....
 412Architectures that care about DMA_ATTR_FOO would check for its
 413presence in their implementations of the mapping and unmapping
 414routines, e.g.:
 416void whizco_dma_map_sg_attrs(struct device *dev, dma_addr_t dma_addr,
 417                             size_t size, enum dma_data_direction dir,
 418                             struct dma_attrs *attrs)
 420        ....
 421        int foo =  dma_get_attr(DMA_ATTR_FOO, attrs);
 422        ....
 423        if (foo)
 424                /* twizzle the frobnozzle */
 425        ....
 428Part II - Advanced dma_ usage
 431Warning: These pieces of the DMA API should not be used in the
 432majority of cases, since they cater for unlikely corner cases that
 433don't belong in usual drivers.
 435If you don't understand how cache line coherency works between a
 436processor and an I/O device, you should not be using this part of the
 437API at all.
 439void *
 440dma_alloc_noncoherent(struct device *dev, size_t size,
 441                               dma_addr_t *dma_handle, gfp_t flag)
 443Identical to dma_alloc_coherent() except that the platform will
 444choose to return either consistent or non-consistent memory as it sees
 445fit.  By using this API, you are guaranteeing to the platform that you
 446have all the correct and necessary sync points for this memory in the
 447driver should it choose to return non-consistent memory.
 449Note: where the platform can return consistent memory, it will
 450guarantee that the sync points become nops.
 452Warning:  Handling non-consistent memory is a real pain.  You should
 453only ever use this API if you positively know your driver will be
 454required to work on one of the rare (usually non-PCI) architectures
 455that simply cannot make consistent memory.
 458dma_free_noncoherent(struct device *dev, size_t size, void *cpu_addr,
 459                              dma_addr_t dma_handle)
 461Free memory allocated by the nonconsistent API.  All parameters must
 462be identical to those passed in (and returned by
 468Returns the processor cache alignment.  This is the absolute minimum
 469alignment *and* width that you must observe when either mapping
 470memory or doing partial flushes.
 472Notes: This API may return a number *larger* than the actual cache
 473line, but it will guarantee that one or more cache lines fit exactly
 474into the width returned by this call.  It will also always be a power
 475of two for easy alignment.
 478dma_cache_sync(struct device *dev, void *vaddr, size_t size,
 479               enum dma_data_direction direction)
 481Do a partial sync of memory that was allocated by
 482dma_alloc_noncoherent(), starting at virtual address vaddr and
 483continuing on for size.  Again, you *must* observe the cache line
 484boundaries when doing this.
 487dma_declare_coherent_memory(struct device *dev, dma_addr_t bus_addr,
 488                            dma_addr_t device_addr, size_t size, int
 489                            flags)
 491Declare region of memory to be handed out by dma_alloc_coherent when
 492it's asked for coherent memory for this device.
 494bus_addr is the physical address to which the memory is currently
 495assigned in the bus responding region (this will be used by the
 496platform to perform the mapping).
 498device_addr is the physical address the device needs to be programmed
 499with actually to address this memory (this will be handed out as the
 500dma_addr_t in dma_alloc_coherent()).
 502size is the size of the area (must be multiples of PAGE_SIZE).
 504flags can be or'd together and are:
 506DMA_MEMORY_MAP - request that the memory returned from
 507dma_alloc_coherent() be directly writable.
 509DMA_MEMORY_IO - request that the memory returned from
 510dma_alloc_coherent() be addressable using read/write/memcpy_toio etc.
 512One or both of these flags must be present.
 514DMA_MEMORY_INCLUDES_CHILDREN - make the declared memory be allocated by
 515dma_alloc_coherent of any child devices of this one (for memory residing
 516on a bridge).
 518DMA_MEMORY_EXCLUSIVE - only allocate memory from the declared regions. 
 519Do not allow dma_alloc_coherent() to fall back to system memory when
 520it's out of memory in the declared region.
 522The return value will be either DMA_MEMORY_MAP or DMA_MEMORY_IO and
 523must correspond to a passed in flag (i.e. no returning DMA_MEMORY_IO
 524if only DMA_MEMORY_MAP were passed in) for success or zero for
 527Note, for DMA_MEMORY_IO returns, all subsequent memory returned by
 528dma_alloc_coherent() may no longer be accessed directly, but instead
 529must be accessed using the correct bus functions.  If your driver
 530isn't prepared to handle this contingency, it should not specify
 531DMA_MEMORY_IO in the input flags.
 533As a simplification for the platforms, only *one* such region of
 534memory may be declared per device.
 536For reasons of efficiency, most platforms choose to track the declared
 537region only at the granularity of a page.  For smaller allocations,
 538you should use the dma_pool() API.
 541dma_release_declared_memory(struct device *dev)
 543Remove the memory region previously declared from the system.  This
 544API performs *no* in-use checking for this region and will return
 545unconditionally having removed all the required structures.  It is the
 546driver's job to ensure that no parts of this memory region are
 547currently in use.
 549void *
 550dma_mark_declared_memory_occupied(struct device *dev,
 551                                  dma_addr_t device_addr, size_t size)
 553This is used to occupy specific regions of the declared space
 554(dma_alloc_coherent() will hand out the first free region it finds).
 556device_addr is the *device* address of the region requested.
 558size is the size (and should be a page-sized multiple).
 560The return value will be either a pointer to the processor virtual
 561address of the memory, or an error (via PTR_ERR()) if any part of the
 562region is occupied.
 564Part III - Debug drivers use of the DMA-API
 567The DMA-API as described above as some constraints. DMA addresses must be
 568released with the corresponding function with the same size for example. With
 569the advent of hardware IOMMUs it becomes more and more important that drivers
 570do not violate those constraints. In the worst case such a violation can
 571result in data corruption up to destroyed filesystems.
 573To debug drivers and find bugs in the usage of the DMA-API checking code can
 574be compiled into the kernel which will tell the developer about those
 575violations. If your architecture supports it you can select the "Enable
 576debugging of DMA-API usage" option in your kernel configuration. Enabling this
 577option has a performance impact. Do not enable it in production kernels.
 579If you boot the resulting kernel will contain code which does some bookkeeping
 580about what DMA memory was allocated for which device. If this code detects an
 581error it prints a warning message with some details into your kernel log. An
 582example warning message may look like this:
 584------------[ cut here ]------------
 585WARNING: at /data2/repos/linux-2.6-iommu/lib/dma-debug.c:448
 586        check_unmap+0x203/0x490()
 587Hardware name:
 588forcedeth 0000:00:08.0: DMA-API: device driver frees DMA memory with wrong
 589        function [device address=0x00000000640444be] [size=66 bytes] [mapped as
 590single] [unmapped as page]
 591Modules linked in: nfsd exportfs bridge stp llc r8169
 592Pid: 0, comm: swapper Tainted: G        W  2.6.28-dmatest-09289-g8bb99c0 #1
 593Call Trace:
 594 <IRQ>  [<ffffffff80240b22>] warn_slowpath+0xf2/0x130
 595 [<ffffffff80647b70>] _spin_unlock+0x10/0x30
 596 [<ffffffff80537e75>] usb_hcd_link_urb_to_ep+0x75/0xc0
 597 [<ffffffff80647c22>] _spin_unlock_irqrestore+0x12/0x40
 598 [<ffffffff8055347f>] ohci_urb_enqueue+0x19f/0x7c0
 599 [<ffffffff80252f96>] queue_work+0x56/0x60
 600 [<ffffffff80237e10>] enqueue_task_fair+0x20/0x50
 601 [<ffffffff80539279>] usb_hcd_submit_urb+0x379/0xbc0
 602 [<ffffffff803b78c3>] cpumask_next_and+0x23/0x40
 603 [<ffffffff80235177>] find_busiest_group+0x207/0x8a0
 604 [<ffffffff8064784f>] _spin_lock_irqsave+0x1f/0x50
 605 [<ffffffff803c7ea3>] check_unmap+0x203/0x490
 606 [<ffffffff803c8259>] debug_dma_unmap_page+0x49/0x50
 607 [<ffffffff80485f26>] nv_tx_done_optimized+0xc6/0x2c0
 608 [<ffffffff80486c13>] nv_nic_irq_optimized+0x73/0x2b0
 609 [<ffffffff8026df84>] handle_IRQ_event+0x34/0x70
 610 [<ffffffff8026ffe9>] handle_edge_irq+0xc9/0x150
 611 [<ffffffff8020e3ab>] do_IRQ+0xcb/0x1c0
 612 [<ffffffff8020c093>] ret_from_intr+0x0/0xa
 613 <EOI> <4>---[ end trace f6435a98e2a38c0e ]---
 615The driver developer can find the driver and the device including a stacktrace
 616of the DMA-API call which caused this warning.
 618Per default only the first error will result in a warning message. All other
 619errors will only silently counted. This limitation exist to prevent the code
 620from flooding your kernel log. To support debugging a device driver this can
 621be disabled via debugfs. See the debugfs interface documentation below for
 624The debugfs directory for the DMA-API debugging code is called dma-api/. In
 625this directory the following files can currently be found:
 627        dma-api/all_errors      This file contains a numeric value. If this
 628                                value is not equal to zero the debugging code
 629                                will print a warning for every error it finds
 630                                into the kernel log. Be careful with this
 631                                option, as it can easily flood your logs.
 633        dma-api/disabled        This read-only file contains the character 'Y'
 634                                if the debugging code is disabled. This can
 635                                happen when it runs out of memory or if it was
 636                                disabled at boot time
 638        dma-api/error_count     This file is read-only and shows the total
 639                                numbers of errors found.
 641        dma-api/num_errors      The number in this file shows how many
 642                                warnings will be printed to the kernel log
 643                                before it stops. This number is initialized to
 644                                one at system boot and be set by writing into
 645                                this file
 647        dma-api/min_free_entries
 648                                This read-only file can be read to get the
 649                                minimum number of free dma_debug_entries the
 650                                allocator has ever seen. If this value goes
 651                                down to zero the code will disable itself
 652                                because it is not longer reliable.
 654        dma-api/num_free_entries
 655                                The current number of free dma_debug_entries
 656                                in the allocator.
 658        dma-api/driver-filter
 659                                You can write a name of a driver into this file
 660                                to limit the debug output to requests from that
 661                                particular driver. Write an empty string to
 662                                that file to disable the filter and see
 663                                all errors again.
 665If you have this code compiled into your kernel it will be enabled by default.
 666If you want to boot without the bookkeeping anyway you can provide
 667'dma_debug=off' as a boot parameter. This will disable DMA-API debugging.
 668Notice that you can not enable it again at runtime. You have to reboot to do
 671If you want to see debug messages only for a special device driver you can
 672specify the dma_debug_driver=<drivername> parameter. This will enable the
 673driver filter at boot time. The debug code will only print errors for that
 674driver afterwards. This filter can be disabled or changed later using debugfs.
 676When the code disables itself at runtime this is most likely because it ran
 677out of dma_debug_entries. These entries are preallocated at boot. The number
 678of preallocated entries is defined per architecture. If it is too low for you
 679boot with 'dma_debug_entries=<your_desired_number>' to overwrite the
 680architectural default.
 682void debug_dmap_mapping_error(struct device *dev, dma_addr_t dma_addr);
 684dma-debug interface debug_dma_mapping_error() to debug drivers that fail
 685to check dma mapping errors on addresses returned by dma_map_single() and
 686dma_map_page() interfaces. This interface clears a flag set by
 687debug_dma_map_page() to indicate that dma_mapping_error() has been called by
 688the driver. When driver does unmap, debug_dma_unmap() checks the flag and if
 689this flag is still set, prints warning message that includes call trace that
 690leads up to the unmap. This interface can be called from dma_mapping_error()
 691routines to enable dma mapping error check debugging.