1                    DMA Buffer Sharing API Guide
   2                    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
   4                            Sumit Semwal
   5                <sumit dot semwal at linaro dot org>
   6                 <sumit dot semwal at ti dot com>
   8This document serves as a guide to device-driver writers on what is the dma-buf
   9buffer sharing API, how to use it for exporting and using shared buffers.
  11Any device driver which wishes to be a part of DMA buffer sharing, can do so as
  12either the 'exporter' of buffers, or the 'user' of buffers.
  14Say a driver A wants to use buffers created by driver B, then we call B as the
  15exporter, and A as buffer-user.
  17The exporter
  18- implements and manages operations[1] for the buffer
  19- allows other users to share the buffer by using dma_buf sharing APIs,
  20- manages the details of buffer allocation,
  21- decides about the actual backing storage where this allocation happens,
  22- takes care of any migration of scatterlist - for all (shared) users of this
  23   buffer,
  25The buffer-user
  26- is one of (many) sharing users of the buffer.
  27- doesn't need to worry about how the buffer is allocated, or where.
  28- needs a mechanism to get access to the scatterlist that makes up this buffer
  29   in memory, mapped into its own address space, so it can access the same area
  30   of memory.
  32dma-buf operations for device dma only
  35The dma_buf buffer sharing API usage contains the following steps:
  371. Exporter announces that it wishes to export a buffer
  382. Userspace gets the file descriptor associated with the exported buffer, and
  39   passes it around to potential buffer-users based on use case
  403. Each buffer-user 'connects' itself to the buffer
  414. When needed, buffer-user requests access to the buffer from exporter
  425. When finished with its use, the buffer-user notifies end-of-DMA to exporter
  436. when buffer-user is done using this buffer completely, it 'disconnects'
  44   itself from the buffer.
  471. Exporter's announcement of buffer export
  49   The buffer exporter announces its wish to export a buffer. In this, it
  50   connects its own private buffer data, provides implementation for operations
  51   that can be performed on the exported dma_buf, and flags for the file
  52   associated with this buffer.
  54   Interface:
  55      struct dma_buf *dma_buf_export(void *priv, struct dma_buf_ops *ops,
  56                                     size_t size, int flags)
  58   If this succeeds, dma_buf_export allocates a dma_buf structure, and returns a
  59   pointer to the same. It also associates an anonymous file with this buffer,
  60   so it can be exported. On failure to allocate the dma_buf object, it returns
  61   NULL.
  632. Userspace gets a handle to pass around to potential buffer-users
  65   Userspace entity requests for a file-descriptor (fd) which is a handle to the
  66   anonymous file associated with the buffer. It can then share the fd with other
  67   drivers and/or processes.
  69   Interface:
  70      int dma_buf_fd(struct dma_buf *dmabuf)
  72   This API installs an fd for the anonymous file associated with this buffer;
  73   returns either 'fd', or error.
  753. Each buffer-user 'connects' itself to the buffer
  77   Each buffer-user now gets a reference to the buffer, using the fd passed to
  78   it.
  80   Interface:
  81      struct dma_buf *dma_buf_get(int fd)
  83   This API will return a reference to the dma_buf, and increment refcount for
  84   it.
  86   After this, the buffer-user needs to attach its device with the buffer, which
  87   helps the exporter to know of device buffer constraints.
  89   Interface:
  90      struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
  91                                                struct device *dev)
  93   This API returns reference to an attachment structure, which is then used
  94   for scatterlist operations. It will optionally call the 'attach' dma_buf
  95   operation, if provided by the exporter.
  97   The dma-buf sharing framework does the bookkeeping bits related to managing
  98   the list of all attachments to a buffer.
 100Until this stage, the buffer-exporter has the option to choose not to actually
 101allocate the backing storage for this buffer, but wait for the first buffer-user
 102to request use of buffer for allocation.
 1054. When needed, buffer-user requests access to the buffer
 107   Whenever a buffer-user wants to use the buffer for any DMA, it asks for
 108   access to the buffer using dma_buf_map_attachment API. At least one attach to
 109   the buffer must have happened before map_dma_buf can be called.
 111   Interface:
 112      struct sg_table * dma_buf_map_attachment(struct dma_buf_attachment *,
 113                                         enum dma_data_direction);
 115   This is a wrapper to dma_buf->ops->map_dma_buf operation, which hides the
 116   "dma_buf->ops->" indirection from the users of this interface.
 118   In struct dma_buf_ops, map_dma_buf is defined as
 119      struct sg_table * (*map_dma_buf)(struct dma_buf_attachment *,
 120                                                enum dma_data_direction);
 122   It is one of the buffer operations that must be implemented by the exporter.
 123   It should return the sg_table containing scatterlist for this buffer, mapped
 124   into caller's address space.
 126   If this is being called for the first time, the exporter can now choose to
 127   scan through the list of attachments for this buffer, collate the requirements
 128   of the attached devices, and choose an appropriate backing storage for the
 129   buffer.
 131   Based on enum dma_data_direction, it might be possible to have multiple users
 132   accessing at the same time (for reading, maybe), or any other kind of sharing
 133   that the exporter might wish to make available to buffer-users.
 135   map_dma_buf() operation can return -EINTR if it is interrupted by a signal.
 1385. When finished, the buffer-user notifies end-of-DMA to exporter
 140   Once the DMA for the current buffer-user is over, it signals 'end-of-DMA' to
 141   the exporter using the dma_buf_unmap_attachment API.
 143   Interface:
 144      void dma_buf_unmap_attachment(struct dma_buf_attachment *,
 145                                    struct sg_table *);
 147   This is a wrapper to dma_buf->ops->unmap_dma_buf() operation, which hides the
 148   "dma_buf->ops->" indirection from the users of this interface.
 150   In struct dma_buf_ops, unmap_dma_buf is defined as
 151      void (*unmap_dma_buf)(struct dma_buf_attachment *, struct sg_table *);
 153   unmap_dma_buf signifies the end-of-DMA for the attachment provided. Like
 154   map_dma_buf, this API also must be implemented by the exporter.
 1576. when buffer-user is done using this buffer, it 'disconnects' itself from the
 158   buffer.
 160   After the buffer-user has no more interest in using this buffer, it should
 161   disconnect itself from the buffer:
 163   - it first detaches itself from the buffer.
 165   Interface:
 166      void dma_buf_detach(struct dma_buf *dmabuf,
 167                          struct dma_buf_attachment *dmabuf_attach);
 169   This API removes the attachment from the list in dmabuf, and optionally calls
 170   dma_buf->ops->detach(), if provided by exporter, for any housekeeping bits.
 172   - Then, the buffer-user returns the buffer reference to exporter.
 174   Interface:
 175     void dma_buf_put(struct dma_buf *dmabuf);
 177   This API then reduces the refcount for this buffer.
 179   If, as a result of this call, the refcount becomes 0, the 'release' file
 180   operation related to this fd is called. It calls the dmabuf->ops->release()
 181   operation in turn, and frees the memory allocated for dmabuf when exported.
 184- Importance of attach-detach and {map,unmap}_dma_buf operation pairs
 185   The attach-detach calls allow the exporter to figure out backing-storage
 186   constraints for the currently-interested devices. This allows preferential
 187   allocation, and/or migration of pages across different types of storage
 188   available, if possible.
 190   Bracketing of DMA access with {map,unmap}_dma_buf operations is essential
 191   to allow just-in-time backing of storage, and migration mid-way through a
 192   use-case.
 194- Migration of backing storage if needed
 195   If after
 196   - at least one map_dma_buf has happened,
 197   - and the backing storage has been allocated for this buffer,
 198   another new buffer-user intends to attach itself to this buffer, it might
 199   be allowed, if possible for the exporter.
 201   In case it is allowed by the exporter:
 202    if the new buffer-user has stricter 'backing-storage constraints', and the
 203    exporter can handle these constraints, the exporter can just stall on the
 204    map_dma_buf until all outstanding access is completed (as signalled by
 205    unmap_dma_buf).
 206    Once all users have finished accessing and have unmapped this buffer, the
 207    exporter could potentially move the buffer to the stricter backing-storage,
 208    and then allow further {map,unmap}_dma_buf operations from any buffer-user
 209    from the migrated backing-storage.
 211   If the exporter cannot fulfil the backing-storage constraints of the new
 212   buffer-user device as requested, dma_buf_attach() would return an error to
 213   denote non-compatibility of the new buffer-sharing request with the current
 214   buffer.
 216   If the exporter chooses not to allow an attach() operation once a
 217   map_dma_buf() API has been called, it simply returns an error.
 219Kernel cpu access to a dma-buf buffer object
 222The motivation to allow cpu access from the kernel to a dma-buf object from the
 223importers side are:
 224- fallback operations, e.g. if the devices is connected to a usb bus and the
 225  kernel needs to shuffle the data around first before sending it away.
 226- full transparency for existing users on the importer side, i.e. userspace
 227  should not notice the difference between a normal object from that subsystem
 228  and an imported one backed by a dma-buf. This is really important for drm
 229  opengl drivers that expect to still use all the existing upload/download
 230  paths.
 232Access to a dma_buf from the kernel context involves three steps:
 2341. Prepare access, which invalidate any necessary caches and make the object
 235   available for cpu access.
 2362. Access the object page-by-page with the dma_buf map apis
 2373. Finish access, which will flush any necessary cpu caches and free reserved
 238   resources.
 2401. Prepare access
 242   Before an importer can access a dma_buf object with the cpu from the kernel
 243   context, it needs to notify the exporter of the access that is about to
 244   happen.
 246   Interface:
 247      int dma_buf_begin_cpu_access(struct dma_buf *dmabuf,
 248                                   size_t start, size_t len,
 249                                   enum dma_data_direction direction)
 251   This allows the exporter to ensure that the memory is actually available for
 252   cpu access - the exporter might need to allocate or swap-in and pin the
 253   backing storage. The exporter also needs to ensure that cpu access is
 254   coherent for the given range and access direction. The range and access
 255   direction can be used by the exporter to optimize the cache flushing, i.e.
 256   access outside of the range or with a different direction (read instead of
 257   write) might return stale or even bogus data (e.g. when the exporter needs to
 258   copy the data to temporary storage).
 260   This step might fail, e.g. in oom conditions.
 2622. Accessing the buffer
 264   To support dma_buf objects residing in highmem cpu access is page-based using
 265   an api similar to kmap. Accessing a dma_buf is done in aligned chunks of
 266   PAGE_SIZE size. Before accessing a chunk it needs to be mapped, which returns
 267   a pointer in kernel virtual address space. Afterwards the chunk needs to be
 268   unmapped again. There is no limit on how often a given chunk can be mapped
 269   and unmapped, i.e. the importer does not need to call begin_cpu_access again
 270   before mapping the same chunk again.
 272   Interfaces:
 273      void *dma_buf_kmap(struct dma_buf *, unsigned long);
 274      void dma_buf_kunmap(struct dma_buf *, unsigned long, void *);
 276   There are also atomic variants of these interfaces. Like for kmap they
 277   facilitate non-blocking fast-paths. Neither the importer nor the exporter (in
 278   the callback) is allowed to block when using these.
 280   Interfaces:
 281      void *dma_buf_kmap_atomic(struct dma_buf *, unsigned long);
 282      void dma_buf_kunmap_atomic(struct dma_buf *, unsigned long, void *);
 284   For importers all the restrictions of using kmap apply, like the limited
 285   supply of kmap_atomic slots. Hence an importer shall only hold onto at most 2
 286   atomic dma_buf kmaps at the same time (in any given process context).
 288   dma_buf kmap calls outside of the range specified in begin_cpu_access are
 289   undefined. If the range is not PAGE_SIZE aligned, kmap needs to succeed on
 290   the partial chunks at the beginning and end but may return stale or bogus
 291   data outside of the range (in these partial chunks).
 293   Note that these calls need to always succeed. The exporter needs to complete
 294   any preparations that might fail in begin_cpu_access.
 296   For some cases the overhead of kmap can be too high, a vmap interface
 297   is introduced. This interface should be used very carefully, as vmalloc
 298   space is a limited resources on many architectures.
 300   Interfaces:
 301      void *dma_buf_vmap(struct dma_buf *dmabuf)
 302      void dma_buf_vunmap(struct dma_buf *dmabuf, void *vaddr)
 304   The vmap call can fail if there is no vmap support in the exporter, or if it
 305   runs out of vmalloc space. Fallback to kmap should be implemented.
 3073. Finish access
 309   When the importer is done accessing the range specified in begin_cpu_access,
 310   it needs to announce this to the exporter (to facilitate cache flushing and
 311   unpinning of any pinned resources). The result of of any dma_buf kmap calls
 312   after end_cpu_access is undefined.
 314   Interface:
 315      void dma_buf_end_cpu_access(struct dma_buf *dma_buf,
 316                                  size_t start, size_t len,
 317                                  enum dma_data_direction dir);
 320Direct Userspace Access/mmap Support
 323Being able to mmap an export dma-buf buffer object has 2 main use-cases:
 324- CPU fallback processing in a pipeline and
 325- supporting existing mmap interfaces in importers.
 3271. CPU fallback processing in a pipeline
 329   In many processing pipelines it is sometimes required that the cpu can access
 330   the data in a dma-buf (e.g. for thumbnail creation, snapshots, ...). To avoid
 331   the need to handle this specially in userspace frameworks for buffer sharing
 332   it's ideal if the dma_buf fd itself can be used to access the backing storage
 333   from userspace using mmap.
 335   Furthermore Android's ION framework already supports this (and is otherwise
 336   rather similar to dma-buf from a userspace consumer side with using fds as
 337   handles, too). So it's beneficial to support this in a similar fashion on
 338   dma-buf to have a good transition path for existing Android userspace.
 340   No special interfaces, userspace simply calls mmap on the dma-buf fd.
 3422. Supporting existing mmap interfaces in exporters
 344   Similar to the motivation for kernel cpu access it is again important that
 345   the userspace code of a given importing subsystem can use the same interfaces
 346   with a imported dma-buf buffer object as with a native buffer object. This is
 347   especially important for drm where the userspace part of contemporary OpenGL,
 348   X, and other drivers is huge, and reworking them to use a different way to
 349   mmap a buffer rather invasive.
 351   The assumption in the current dma-buf interfaces is that redirecting the
 352   initial mmap is all that's needed. A survey of some of the existing
 353   subsystems shows that no driver seems to do any nefarious thing like syncing
 354   up with outstanding asynchronous processing on the device or allocating
 355   special resources at fault time. So hopefully this is good enough, since
 356   adding interfaces to intercept pagefaults and allow pte shootdowns would
 357   increase the complexity quite a bit.
 359   Interface:
 360      int dma_buf_mmap(struct dma_buf *, struct vm_area_struct *,
 361                       unsigned long);
 363   If the importing subsystem simply provides a special-purpose mmap call to set
 364   up a mapping in userspace, calling do_mmap with dma_buf->file will equally
 365   achieve that for a dma-buf object.
 3673. Implementation notes for exporters
 369   Because dma-buf buffers have invariant size over their lifetime, the dma-buf
 370   core checks whether a vma is too large and rejects such mappings. The
 371   exporter hence does not need to duplicate this check.
 373   Because existing importing subsystems might presume coherent mappings for
 374   userspace, the exporter needs to set up a coherent mapping. If that's not
 375   possible, it needs to fake coherency by manually shooting down ptes when
 376   leaving the cpu domain and flushing caches at fault time. Note that all the
 377   dma_buf files share the same anon inode, hence the exporter needs to replace
 378   the dma_buf file stored in vma->vm_file with it's own if pte shootdown is
 379   required. This is because the kernel uses the underlying inode's address_space
 380   for vma tracking (and hence pte tracking at shootdown time with
 381   unmap_mapping_range).
 383   If the above shootdown dance turns out to be too expensive in certain
 384   scenarios, we can extend dma-buf with a more explicit cache tracking scheme
 385   for userspace mappings. But the current assumption is that using mmap is
 386   always a slower path, so some inefficiencies should be acceptable.
 388   Exporters that shoot down mappings (for any reasons) shall not do any
 389   synchronization at fault time with outstanding device operations.
 390   Synchronization is an orthogonal issue to sharing the backing storage of a
 391   buffer and hence should not be handled by dma-buf itself. This is explicitly
 392   mentioned here because many people seem to want something like this, but if
 393   different exporters handle this differently, buffer sharing can fail in
 394   interesting ways depending upong the exporter (if userspace starts depending
 395   upon this implicit synchronization).
 397Miscellaneous notes
 400- Any exporters or users of the dma-buf buffer sharing framework must have
 401  a 'select DMA_SHARED_BUFFER' in their respective Kconfigs.
 403- In order to avoid fd leaks on exec, the FD_CLOEXEC flag must be set
 404  on the file descriptor.  This is not just a resource leak, but a
 405  potential security hole.  It could give the newly exec'd application
 406  access to buffers, via the leaked fd, to which it should otherwise
 407  not be permitted access.
 409  The problem with doing this via a separate fcntl() call, versus doing it
 410  atomically when the fd is created, is that this is inherently racy in a
 411  multi-threaded app[3].  The issue is made worse when it is library code
 412  opening/creating the file descriptor, as the application may not even be
 413  aware of the fd's.
 415  To avoid this problem, userspace must have a way to request O_CLOEXEC
 416  flag be set when the dma-buf fd is created.  So any API provided by
 417  the exporting driver to create a dmabuf fd must provide a way to let
 418  userspace control setting of O_CLOEXEC flag passed in to dma_buf_fd().
 420- If an exporter needs to manually flush caches and hence needs to fake
 421  coherency for mmap support, it needs to be able to zap all the ptes pointing
 422  at the backing storage. Now linux mm needs a struct address_space associated
 423  with the struct file stored in vma->vm_file to do that with the function
 424  unmap_mapping_range. But the dma_buf framework only backs every dma_buf fd
 425  with the anon_file struct file, i.e. all dma_bufs share the same file.
 427  Hence exporters need to setup their own file (and address_space) association
 428  by setting vma->vm_file and adjusting vma->vm_pgoff in the dma_buf mmap
 429  callback. In the specific case of a gem driver the exporter could use the
 430  shmem file already provided by gem (and set vm_pgoff = 0). Exporters can then
 431  zap ptes by unmapping the corresponding range of the struct address_space
 432  associated with their own file.
 435[1] struct dma_buf_ops in include/linux/dma-buf.h
 436[2] All interfaces mentioned above defined in include/linux/dma-buf.h