Root/Documentation/dma-buf-sharing.txt

1                    DMA Buffer Sharing API Guide
2                    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
3
4                            Sumit Semwal
5                <sumit dot semwal at linaro dot org>
6                 <sumit dot semwal at ti dot com>
7
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.
10
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.
13
14Say a driver A wants to use buffers created by driver B, then we call B as the
15exporter, and A as buffer-user.
16
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,
24
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.
31
32dma-buf operations for device dma only
33--------------------------------------
34
35The dma_buf buffer sharing API usage contains the following steps:
36
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.
45
46
471. Exporter's announcement of buffer export
48
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.
53
54   Interface:
55      struct dma_buf *dma_buf_export_named(void *priv, struct dma_buf_ops *ops,
56                     size_t size, int flags,
57                     const char *exp_name)
58
59   If this succeeds, dma_buf_export allocates a dma_buf structure, and returns a
60   pointer to the same. It also associates an anonymous file with this buffer,
61   so it can be exported. On failure to allocate the dma_buf object, it returns
62   NULL.
63
64   'exp_name' is the name of exporter - to facilitate information while
65   debugging.
66
67   Exporting modules which do not wish to provide any specific name may use the
68   helper define 'dma_buf_export()', with the same arguments as above, but
69   without the last argument; a __FILE__ pre-processor directive will be
70   inserted in place of 'exp_name' instead.
71
722. Userspace gets a handle to pass around to potential buffer-users
73
74   Userspace entity requests for a file-descriptor (fd) which is a handle to the
75   anonymous file associated with the buffer. It can then share the fd with other
76   drivers and/or processes.
77
78   Interface:
79      int dma_buf_fd(struct dma_buf *dmabuf)
80
81   This API installs an fd for the anonymous file associated with this buffer;
82   returns either 'fd', or error.
83
843. Each buffer-user 'connects' itself to the buffer
85
86   Each buffer-user now gets a reference to the buffer, using the fd passed to
87   it.
88
89   Interface:
90      struct dma_buf *dma_buf_get(int fd)
91
92   This API will return a reference to the dma_buf, and increment refcount for
93   it.
94
95   After this, the buffer-user needs to attach its device with the buffer, which
96   helps the exporter to know of device buffer constraints.
97
98   Interface:
99      struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf,
100                                                struct device *dev)
101
102   This API returns reference to an attachment structure, which is then used
103   for scatterlist operations. It will optionally call the 'attach' dma_buf
104   operation, if provided by the exporter.
105
106   The dma-buf sharing framework does the bookkeeping bits related to managing
107   the list of all attachments to a buffer.
108
109Until this stage, the buffer-exporter has the option to choose not to actually
110allocate the backing storage for this buffer, but wait for the first buffer-user
111to request use of buffer for allocation.
112
113
1144. When needed, buffer-user requests access to the buffer
115
116   Whenever a buffer-user wants to use the buffer for any DMA, it asks for
117   access to the buffer using dma_buf_map_attachment API. At least one attach to
118   the buffer must have happened before map_dma_buf can be called.
119
120   Interface:
121      struct sg_table * dma_buf_map_attachment(struct dma_buf_attachment *,
122                                         enum dma_data_direction);
123
124   This is a wrapper to dma_buf->ops->map_dma_buf operation, which hides the
125   "dma_buf->ops->" indirection from the users of this interface.
126
127   In struct dma_buf_ops, map_dma_buf is defined as
128      struct sg_table * (*map_dma_buf)(struct dma_buf_attachment *,
129                                                enum dma_data_direction);
130
131   It is one of the buffer operations that must be implemented by the exporter.
132   It should return the sg_table containing scatterlist for this buffer, mapped
133   into caller's address space.
134
135   If this is being called for the first time, the exporter can now choose to
136   scan through the list of attachments for this buffer, collate the requirements
137   of the attached devices, and choose an appropriate backing storage for the
138   buffer.
139
140   Based on enum dma_data_direction, it might be possible to have multiple users
141   accessing at the same time (for reading, maybe), or any other kind of sharing
142   that the exporter might wish to make available to buffer-users.
143
144   map_dma_buf() operation can return -EINTR if it is interrupted by a signal.
145
146
1475. When finished, the buffer-user notifies end-of-DMA to exporter
148
149   Once the DMA for the current buffer-user is over, it signals 'end-of-DMA' to
150   the exporter using the dma_buf_unmap_attachment API.
151
152   Interface:
153      void dma_buf_unmap_attachment(struct dma_buf_attachment *,
154                                    struct sg_table *);
155
156   This is a wrapper to dma_buf->ops->unmap_dma_buf() operation, which hides the
157   "dma_buf->ops->" indirection from the users of this interface.
158
159   In struct dma_buf_ops, unmap_dma_buf is defined as
160      void (*unmap_dma_buf)(struct dma_buf_attachment *, struct sg_table *);
161
162   unmap_dma_buf signifies the end-of-DMA for the attachment provided. Like
163   map_dma_buf, this API also must be implemented by the exporter.
164
165
1666. when buffer-user is done using this buffer, it 'disconnects' itself from the
167   buffer.
168
169   After the buffer-user has no more interest in using this buffer, it should
170   disconnect itself from the buffer:
171
172   - it first detaches itself from the buffer.
173
174   Interface:
175      void dma_buf_detach(struct dma_buf *dmabuf,
176                          struct dma_buf_attachment *dmabuf_attach);
177
178   This API removes the attachment from the list in dmabuf, and optionally calls
179   dma_buf->ops->detach(), if provided by exporter, for any housekeeping bits.
180
181   - Then, the buffer-user returns the buffer reference to exporter.
182
183   Interface:
184     void dma_buf_put(struct dma_buf *dmabuf);
185
186   This API then reduces the refcount for this buffer.
187
188   If, as a result of this call, the refcount becomes 0, the 'release' file
189   operation related to this fd is called. It calls the dmabuf->ops->release()
190   operation in turn, and frees the memory allocated for dmabuf when exported.
191
192NOTES:
193- Importance of attach-detach and {map,unmap}_dma_buf operation pairs
194   The attach-detach calls allow the exporter to figure out backing-storage
195   constraints for the currently-interested devices. This allows preferential
196   allocation, and/or migration of pages across different types of storage
197   available, if possible.
198
199   Bracketing of DMA access with {map,unmap}_dma_buf operations is essential
200   to allow just-in-time backing of storage, and migration mid-way through a
201   use-case.
202
203- Migration of backing storage if needed
204   If after
205   - at least one map_dma_buf has happened,
206   - and the backing storage has been allocated for this buffer,
207   another new buffer-user intends to attach itself to this buffer, it might
208   be allowed, if possible for the exporter.
209
210   In case it is allowed by the exporter:
211    if the new buffer-user has stricter 'backing-storage constraints', and the
212    exporter can handle these constraints, the exporter can just stall on the
213    map_dma_buf until all outstanding access is completed (as signalled by
214    unmap_dma_buf).
215    Once all users have finished accessing and have unmapped this buffer, the
216    exporter could potentially move the buffer to the stricter backing-storage,
217    and then allow further {map,unmap}_dma_buf operations from any buffer-user
218    from the migrated backing-storage.
219
220   If the exporter cannot fulfil the backing-storage constraints of the new
221   buffer-user device as requested, dma_buf_attach() would return an error to
222   denote non-compatibility of the new buffer-sharing request with the current
223   buffer.
224
225   If the exporter chooses not to allow an attach() operation once a
226   map_dma_buf() API has been called, it simply returns an error.
227
228Kernel cpu access to a dma-buf buffer object
229--------------------------------------------
230
231The motivation to allow cpu access from the kernel to a dma-buf object from the
232importers side are:
233- fallback operations, e.g. if the devices is connected to a usb bus and the
234  kernel needs to shuffle the data around first before sending it away.
235- full transparency for existing users on the importer side, i.e. userspace
236  should not notice the difference between a normal object from that subsystem
237  and an imported one backed by a dma-buf. This is really important for drm
238  opengl drivers that expect to still use all the existing upload/download
239  paths.
240
241Access to a dma_buf from the kernel context involves three steps:
242
2431. Prepare access, which invalidate any necessary caches and make the object
244   available for cpu access.
2452. Access the object page-by-page with the dma_buf map apis
2463. Finish access, which will flush any necessary cpu caches and free reserved
247   resources.
248
2491. Prepare access
250
251   Before an importer can access a dma_buf object with the cpu from the kernel
252   context, it needs to notify the exporter of the access that is about to
253   happen.
254
255   Interface:
256      int dma_buf_begin_cpu_access(struct dma_buf *dmabuf,
257                   size_t start, size_t len,
258                   enum dma_data_direction direction)
259
260   This allows the exporter to ensure that the memory is actually available for
261   cpu access - the exporter might need to allocate or swap-in and pin the
262   backing storage. The exporter also needs to ensure that cpu access is
263   coherent for the given range and access direction. The range and access
264   direction can be used by the exporter to optimize the cache flushing, i.e.
265   access outside of the range or with a different direction (read instead of
266   write) might return stale or even bogus data (e.g. when the exporter needs to
267   copy the data to temporary storage).
268
269   This step might fail, e.g. in oom conditions.
270
2712. Accessing the buffer
272
273   To support dma_buf objects residing in highmem cpu access is page-based using
274   an api similar to kmap. Accessing a dma_buf is done in aligned chunks of
275   PAGE_SIZE size. Before accessing a chunk it needs to be mapped, which returns
276   a pointer in kernel virtual address space. Afterwards the chunk needs to be
277   unmapped again. There is no limit on how often a given chunk can be mapped
278   and unmapped, i.e. the importer does not need to call begin_cpu_access again
279   before mapping the same chunk again.
280
281   Interfaces:
282      void *dma_buf_kmap(struct dma_buf *, unsigned long);
283      void dma_buf_kunmap(struct dma_buf *, unsigned long, void *);
284
285   There are also atomic variants of these interfaces. Like for kmap they
286   facilitate non-blocking fast-paths. Neither the importer nor the exporter (in
287   the callback) is allowed to block when using these.
288
289   Interfaces:
290      void *dma_buf_kmap_atomic(struct dma_buf *, unsigned long);
291      void dma_buf_kunmap_atomic(struct dma_buf *, unsigned long, void *);
292
293   For importers all the restrictions of using kmap apply, like the limited
294   supply of kmap_atomic slots. Hence an importer shall only hold onto at most 2
295   atomic dma_buf kmaps at the same time (in any given process context).
296
297   dma_buf kmap calls outside of the range specified in begin_cpu_access are
298   undefined. If the range is not PAGE_SIZE aligned, kmap needs to succeed on
299   the partial chunks at the beginning and end but may return stale or bogus
300   data outside of the range (in these partial chunks).
301
302   Note that these calls need to always succeed. The exporter needs to complete
303   any preparations that might fail in begin_cpu_access.
304
305   For some cases the overhead of kmap can be too high, a vmap interface
306   is introduced. This interface should be used very carefully, as vmalloc
307   space is a limited resources on many architectures.
308
309   Interfaces:
310      void *dma_buf_vmap(struct dma_buf *dmabuf)
311      void dma_buf_vunmap(struct dma_buf *dmabuf, void *vaddr)
312
313   The vmap call can fail if there is no vmap support in the exporter, or if it
314   runs out of vmalloc space. Fallback to kmap should be implemented. Note that
315   the dma-buf layer keeps a reference count for all vmap access and calls down
316   into the exporter's vmap function only when no vmapping exists, and only
317   unmaps it once. Protection against concurrent vmap/vunmap calls is provided
318   by taking the dma_buf->lock mutex.
319
3203. Finish access
321
322   When the importer is done accessing the range specified in begin_cpu_access,
323   it needs to announce this to the exporter (to facilitate cache flushing and
324   unpinning of any pinned resources). The result of any dma_buf kmap calls
325   after end_cpu_access is undefined.
326
327   Interface:
328      void dma_buf_end_cpu_access(struct dma_buf *dma_buf,
329                  size_t start, size_t len,
330                  enum dma_data_direction dir);
331
332
333Direct Userspace Access/mmap Support
334------------------------------------
335
336Being able to mmap an export dma-buf buffer object has 2 main use-cases:
337- CPU fallback processing in a pipeline and
338- supporting existing mmap interfaces in importers.
339
3401. CPU fallback processing in a pipeline
341
342   In many processing pipelines it is sometimes required that the cpu can access
343   the data in a dma-buf (e.g. for thumbnail creation, snapshots, ...). To avoid
344   the need to handle this specially in userspace frameworks for buffer sharing
345   it's ideal if the dma_buf fd itself can be used to access the backing storage
346   from userspace using mmap.
347
348   Furthermore Android's ION framework already supports this (and is otherwise
349   rather similar to dma-buf from a userspace consumer side with using fds as
350   handles, too). So it's beneficial to support this in a similar fashion on
351   dma-buf to have a good transition path for existing Android userspace.
352
353   No special interfaces, userspace simply calls mmap on the dma-buf fd.
354
3552. Supporting existing mmap interfaces in exporters
356
357   Similar to the motivation for kernel cpu access it is again important that
358   the userspace code of a given importing subsystem can use the same interfaces
359   with a imported dma-buf buffer object as with a native buffer object. This is
360   especially important for drm where the userspace part of contemporary OpenGL,
361   X, and other drivers is huge, and reworking them to use a different way to
362   mmap a buffer rather invasive.
363
364   The assumption in the current dma-buf interfaces is that redirecting the
365   initial mmap is all that's needed. A survey of some of the existing
366   subsystems shows that no driver seems to do any nefarious thing like syncing
367   up with outstanding asynchronous processing on the device or allocating
368   special resources at fault time. So hopefully this is good enough, since
369   adding interfaces to intercept pagefaults and allow pte shootdowns would
370   increase the complexity quite a bit.
371
372   Interface:
373      int dma_buf_mmap(struct dma_buf *, struct vm_area_struct *,
374               unsigned long);
375
376   If the importing subsystem simply provides a special-purpose mmap call to set
377   up a mapping in userspace, calling do_mmap with dma_buf->file will equally
378   achieve that for a dma-buf object.
379
3803. Implementation notes for exporters
381
382   Because dma-buf buffers have invariant size over their lifetime, the dma-buf
383   core checks whether a vma is too large and rejects such mappings. The
384   exporter hence does not need to duplicate this check.
385
386   Because existing importing subsystems might presume coherent mappings for
387   userspace, the exporter needs to set up a coherent mapping. If that's not
388   possible, it needs to fake coherency by manually shooting down ptes when
389   leaving the cpu domain and flushing caches at fault time. Note that all the
390   dma_buf files share the same anon inode, hence the exporter needs to replace
391   the dma_buf file stored in vma->vm_file with it's own if pte shootdown is
392   required. This is because the kernel uses the underlying inode's address_space
393   for vma tracking (and hence pte tracking at shootdown time with
394   unmap_mapping_range).
395
396   If the above shootdown dance turns out to be too expensive in certain
397   scenarios, we can extend dma-buf with a more explicit cache tracking scheme
398   for userspace mappings. But the current assumption is that using mmap is
399   always a slower path, so some inefficiencies should be acceptable.
400
401   Exporters that shoot down mappings (for any reasons) shall not do any
402   synchronization at fault time with outstanding device operations.
403   Synchronization is an orthogonal issue to sharing the backing storage of a
404   buffer and hence should not be handled by dma-buf itself. This is explicitly
405   mentioned here because many people seem to want something like this, but if
406   different exporters handle this differently, buffer sharing can fail in
407   interesting ways depending upong the exporter (if userspace starts depending
408   upon this implicit synchronization).
409
410Other Interfaces Exposed to Userspace on the dma-buf FD
411------------------------------------------------------
412
413- Since kernel 3.12 the dma-buf FD supports the llseek system call, but only
414  with offset=0 and whence=SEEK_END|SEEK_SET. SEEK_SET is supported to allow
415  the usual size discover pattern size = SEEK_END(0); SEEK_SET(0). Every other
416  llseek operation will report -EINVAL.
417
418  If llseek on dma-buf FDs isn't support the kernel will report -ESPIPE for all
419  cases. Userspace can use this to detect support for discovering the dma-buf
420  size using llseek.
421
422Miscellaneous notes
423-------------------
424
425- Any exporters or users of the dma-buf buffer sharing framework must have
426  a 'select DMA_SHARED_BUFFER' in their respective Kconfigs.
427
428- In order to avoid fd leaks on exec, the FD_CLOEXEC flag must be set
429  on the file descriptor. This is not just a resource leak, but a
430  potential security hole. It could give the newly exec'd application
431  access to buffers, via the leaked fd, to which it should otherwise
432  not be permitted access.
433
434  The problem with doing this via a separate fcntl() call, versus doing it
435  atomically when the fd is created, is that this is inherently racy in a
436  multi-threaded app[3]. The issue is made worse when it is library code
437  opening/creating the file descriptor, as the application may not even be
438  aware of the fd's.
439
440  To avoid this problem, userspace must have a way to request O_CLOEXEC
441  flag be set when the dma-buf fd is created. So any API provided by
442  the exporting driver to create a dmabuf fd must provide a way to let
443  userspace control setting of O_CLOEXEC flag passed in to dma_buf_fd().
444
445- If an exporter needs to manually flush caches and hence needs to fake
446  coherency for mmap support, it needs to be able to zap all the ptes pointing
447  at the backing storage. Now linux mm needs a struct address_space associated
448  with the struct file stored in vma->vm_file to do that with the function
449  unmap_mapping_range. But the dma_buf framework only backs every dma_buf fd
450  with the anon_file struct file, i.e. all dma_bufs share the same file.
451
452  Hence exporters need to setup their own file (and address_space) association
453  by setting vma->vm_file and adjusting vma->vm_pgoff in the dma_buf mmap
454  callback. In the specific case of a gem driver the exporter could use the
455  shmem file already provided by gem (and set vm_pgoff = 0). Exporters can then
456  zap ptes by unmapping the corresponding range of the struct address_space
457  associated with their own file.
458
459References:
460[1] struct dma_buf_ops in include/linux/dma-buf.h
461[2] All interfaces mentioned above defined in include/linux/dma-buf.h
462[3] https://lwn.net/Articles/236486/
463

Archive Download this file



interactive