blob: f5760b140f13f19a3e27a037355e2875ea67006c [file] [log] [blame]
Jani Nikula2fa91d12016-06-21 14:49:02 +03001=====================
2DRM Memory Management
3=====================
4
5Modern Linux systems require large amount of graphics memory to store
6frame buffers, textures, vertices and other graphics-related data. Given
7the very dynamic nature of many of that data, managing graphics memory
8efficiently is thus crucial for the graphics stack and plays a central
9role in the DRM infrastructure.
10
11The DRM core includes two memory managers, namely Translation Table Maps
12(TTM) and Graphics Execution Manager (GEM). TTM was the first DRM memory
13manager to be developed and tried to be a one-size-fits-them all
14solution. It provides a single userspace API to accommodate the need of
15all hardware, supporting both Unified Memory Architecture (UMA) devices
16and devices with dedicated video RAM (i.e. most discrete video cards).
17This resulted in a large, complex piece of code that turned out to be
18hard to use for driver development.
19
20GEM started as an Intel-sponsored project in reaction to TTM's
21complexity. Its design philosophy is completely different: instead of
22providing a solution to every graphics memory-related problems, GEM
23identified common code between drivers and created a support library to
24share it. GEM has simpler initialization and execution requirements than
25TTM, but has no video RAM management capabilities and is thus limited to
26UMA devices.
27
28The Translation Table Manager (TTM)
Daniel Vetter8febdf02016-08-12 22:48:40 +020029===================================
Jani Nikula2fa91d12016-06-21 14:49:02 +030030
31TTM design background and information belongs here.
32
33TTM initialization
Daniel Vetter8febdf02016-08-12 22:48:40 +020034------------------
Jani Nikula2fa91d12016-06-21 14:49:02 +030035
36 **Warning**
Jani Nikula2fa91d12016-06-21 14:49:02 +030037 This section is outdated.
38
Gabriel Krisman Bertazib834ff82016-12-30 12:50:49 +010039Drivers wishing to support TTM must pass a filled :c:type:`ttm_bo_driver
40<ttm_bo_driver>` structure to ttm_bo_device_init, together with an
41initialized global reference to the memory manager. The ttm_bo_driver
42structure contains several fields with function pointers for
43initializing the TTM, allocating and freeing memory, waiting for command
44completion and fence synchronization, and memory migration.
Jani Nikula2fa91d12016-06-21 14:49:02 +030045
Gabriel Krisman Bertazib834ff82016-12-30 12:50:49 +010046The :c:type:`struct drm_global_reference <drm_global_reference>` is made
47up of several fields:
Jani Nikula2fa91d12016-06-21 14:49:02 +030048
Jani Nikula29849a62016-11-03 11:44:23 +020049.. code-block:: c
Jani Nikula2fa91d12016-06-21 14:49:02 +030050
Gabriel Krisman Bertazib834ff82016-12-30 12:50:49 +010051 struct drm_global_reference {
Jani Nikula2fa91d12016-06-21 14:49:02 +030052 enum ttm_global_types global_type;
53 size_t size;
54 void *object;
Gabriel Krisman Bertazib834ff82016-12-30 12:50:49 +010055 int (*init) (struct drm_global_reference *);
56 void (*release) (struct drm_global_reference *);
Jani Nikula2fa91d12016-06-21 14:49:02 +030057 };
58
59
60There should be one global reference structure for your memory manager
61as a whole, and there will be others for each object created by the
62memory manager at runtime. Your global TTM should have a type of
63TTM_GLOBAL_TTM_MEM. The size field for the global object should be
64sizeof(struct ttm_mem_global), and the init and release hooks should
65point at your driver-specific init and release routines, which probably
66eventually call ttm_mem_global_init and ttm_mem_global_release,
67respectively.
68
69Once your global TTM accounting structure is set up and initialized by
70calling ttm_global_item_ref() on it, you need to create a buffer
71object TTM to provide a pool for buffer object allocation by clients and
72the kernel itself. The type of this object should be
73TTM_GLOBAL_TTM_BO, and its size should be sizeof(struct
74ttm_bo_global). Again, driver-specific init and release functions may
75be provided, likely eventually calling ttm_bo_global_init() and
76ttm_bo_global_release(), respectively. Also, like the previous
77object, ttm_global_item_ref() is used to create an initial reference
78count for the TTM, which will call your initialization function.
79
Gabriel Krisman Bertazib834ff82016-12-30 12:50:49 +010080See the radeon_ttm.c file for an example of usage.
81
82.. kernel-doc:: drivers/gpu/drm/drm_global.c
83 :export:
84
85
Jani Nikula2fa91d12016-06-21 14:49:02 +030086The Graphics Execution Manager (GEM)
Daniel Vetter8febdf02016-08-12 22:48:40 +020087====================================
Jani Nikula2fa91d12016-06-21 14:49:02 +030088
89The GEM design approach has resulted in a memory manager that doesn't
90provide full coverage of all (or even all common) use cases in its
91userspace or kernel API. GEM exposes a set of standard memory-related
92operations to userspace and a set of helper functions to drivers, and
93let drivers implement hardware-specific operations with their own
94private API.
95
96The GEM userspace API is described in the `GEM - the Graphics Execution
97Manager <http://lwn.net/Articles/283798/>`__ article on LWN. While
98slightly outdated, the document provides a good overview of the GEM API
99principles. Buffer allocation and read and write operations, described
100as part of the common GEM API, are currently implemented using
101driver-specific ioctls.
102
103GEM is data-agnostic. It manages abstract buffer objects without knowing
104what individual buffers contain. APIs that require knowledge of buffer
105contents or purpose, such as buffer allocation or synchronization
106primitives, are thus outside of the scope of GEM and must be implemented
107using driver-specific ioctls.
108
109On a fundamental level, GEM involves several operations:
110
111- Memory allocation and freeing
112- Command execution
113- Aperture management at command execution time
114
115Buffer object allocation is relatively straightforward and largely
116provided by Linux's shmem layer, which provides memory to back each
117object.
118
119Device-specific operations, such as command execution, pinning, buffer
120read & write, mapping, and domain ownership transfers are left to
121driver-specific ioctls.
122
123GEM Initialization
Daniel Vetter8febdf02016-08-12 22:48:40 +0200124------------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300125
126Drivers that use GEM must set the DRIVER_GEM bit in the struct
127:c:type:`struct drm_driver <drm_driver>` driver_features
128field. The DRM core will then automatically initialize the GEM core
129before calling the load operation. Behind the scene, this will create a
130DRM Memory Manager object which provides an address space pool for
131object allocation.
132
133In a KMS configuration, drivers need to allocate and initialize a
134command ring buffer following core GEM initialization if required by the
135hardware. UMA devices usually have what is called a "stolen" memory
136region, which provides space for the initial framebuffer and large,
137contiguous memory regions required by the device. This space is
138typically not managed by GEM, and must be initialized separately into
139its own DRM MM object.
140
141GEM Objects Creation
Daniel Vetter8febdf02016-08-12 22:48:40 +0200142--------------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300143
144GEM splits creation of GEM objects and allocation of the memory that
145backs them in two distinct operations.
146
147GEM objects are represented by an instance of struct :c:type:`struct
148drm_gem_object <drm_gem_object>`. Drivers usually need to
149extend GEM objects with private information and thus create a
150driver-specific GEM object structure type that embeds an instance of
151struct :c:type:`struct drm_gem_object <drm_gem_object>`.
152
153To create a GEM object, a driver allocates memory for an instance of its
154specific GEM object type and initializes the embedded struct
155:c:type:`struct drm_gem_object <drm_gem_object>` with a call
156to :c:func:`drm_gem_object_init()`. The function takes a pointer
157to the DRM device, a pointer to the GEM object and the buffer object
158size in bytes.
159
160GEM uses shmem to allocate anonymous pageable memory.
161:c:func:`drm_gem_object_init()` will create an shmfs file of the
162requested size and store it into the struct :c:type:`struct
163drm_gem_object <drm_gem_object>` filp field. The memory is
164used as either main storage for the object when the graphics hardware
165uses system memory directly or as a backing store otherwise.
166
167Drivers are responsible for the actual physical pages allocation by
168calling :c:func:`shmem_read_mapping_page_gfp()` for each page.
169Note that they can decide to allocate pages when initializing the GEM
170object, or to delay allocation until the memory is needed (for instance
171when a page fault occurs as a result of a userspace memory access or
172when the driver needs to start a DMA transfer involving the memory).
173
174Anonymous pageable memory allocation is not always desired, for instance
175when the hardware requires physically contiguous system memory as is
176often the case in embedded devices. Drivers can create GEM objects with
177no shmfs backing (called private GEM objects) by initializing them with
178a call to :c:func:`drm_gem_private_object_init()` instead of
179:c:func:`drm_gem_object_init()`. Storage for private GEM objects
180must be managed by drivers.
181
182GEM Objects Lifetime
Daniel Vetter8febdf02016-08-12 22:48:40 +0200183--------------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300184
185All GEM objects are reference-counted by the GEM core. References can be
186acquired and release by :c:func:`calling
187drm_gem_object_reference()` and
188:c:func:`drm_gem_object_unreference()` respectively. The caller
189must hold the :c:type:`struct drm_device <drm_device>`
190struct_mutex lock when calling
191:c:func:`drm_gem_object_reference()`. As a convenience, GEM
192provides :c:func:`drm_gem_object_unreference_unlocked()`
193functions that can be called without holding the lock.
194
195When the last reference to a GEM object is released the GEM core calls
196the :c:type:`struct drm_driver <drm_driver>` gem_free_object
197operation. That operation is mandatory for GEM-enabled drivers and must
198free the GEM object and all associated resources.
199
200void (\*gem_free_object) (struct drm_gem_object \*obj); Drivers are
201responsible for freeing all GEM object resources. This includes the
202resources created by the GEM core, which need to be released with
203:c:func:`drm_gem_object_release()`.
204
205GEM Objects Naming
Daniel Vetter8febdf02016-08-12 22:48:40 +0200206------------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300207
208Communication between userspace and the kernel refers to GEM objects
209using local handles, global names or, more recently, file descriptors.
210All of those are 32-bit integer values; the usual Linux kernel limits
211apply to the file descriptors.
212
213GEM handles are local to a DRM file. Applications get a handle to a GEM
214object through a driver-specific ioctl, and can use that handle to refer
215to the GEM object in other standard or driver-specific ioctls. Closing a
216DRM file handle frees all its GEM handles and dereferences the
217associated GEM objects.
218
219To create a handle for a GEM object drivers call
220:c:func:`drm_gem_handle_create()`. The function takes a pointer
221to the DRM file and the GEM object and returns a locally unique handle.
222When the handle is no longer needed drivers delete it with a call to
223:c:func:`drm_gem_handle_delete()`. Finally the GEM object
224associated with a handle can be retrieved by a call to
225:c:func:`drm_gem_object_lookup()`.
226
227Handles don't take ownership of GEM objects, they only take a reference
228to the object that will be dropped when the handle is destroyed. To
229avoid leaking GEM objects, drivers must make sure they drop the
230reference(s) they own (such as the initial reference taken at object
231creation time) as appropriate, without any special consideration for the
232handle. For example, in the particular case of combined GEM object and
233handle creation in the implementation of the dumb_create operation,
234drivers must drop the initial reference to the GEM object before
235returning the handle.
236
237GEM names are similar in purpose to handles but are not local to DRM
238files. They can be passed between processes to reference a GEM object
239globally. Names can't be used directly to refer to objects in the DRM
240API, applications must convert handles to names and names to handles
241using the DRM_IOCTL_GEM_FLINK and DRM_IOCTL_GEM_OPEN ioctls
242respectively. The conversion is handled by the DRM core without any
243driver-specific support.
244
245GEM also supports buffer sharing with dma-buf file descriptors through
246PRIME. GEM-based drivers must use the provided helpers functions to
247implement the exporting and importing correctly. See ?. Since sharing
248file descriptors is inherently more secure than the easily guessable and
249global GEM names it is the preferred buffer sharing mechanism. Sharing
250buffers through GEM names is only supported for legacy userspace.
251Furthermore PRIME also allows cross-device buffer sharing since it is
252based on dma-bufs.
253
254GEM Objects Mapping
Daniel Vetter8febdf02016-08-12 22:48:40 +0200255-------------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300256
257Because mapping operations are fairly heavyweight GEM favours
258read/write-like access to buffers, implemented through driver-specific
259ioctls, over mapping buffers to userspace. However, when random access
260to the buffer is needed (to perform software rendering for instance),
261direct access to the object can be more efficient.
262
263The mmap system call can't be used directly to map GEM objects, as they
264don't have their own file handle. Two alternative methods currently
265co-exist to map GEM objects to userspace. The first method uses a
266driver-specific ioctl to perform the mapping operation, calling
267:c:func:`do_mmap()` under the hood. This is often considered
268dubious, seems to be discouraged for new GEM-enabled drivers, and will
269thus not be described here.
270
271The second method uses the mmap system call on the DRM file handle. void
272\*mmap(void \*addr, size_t length, int prot, int flags, int fd, off_t
273offset); DRM identifies the GEM object to be mapped by a fake offset
274passed through the mmap offset argument. Prior to being mapped, a GEM
275object must thus be associated with a fake offset. To do so, drivers
276must call :c:func:`drm_gem_create_mmap_offset()` on the object.
277
278Once allocated, the fake offset value must be passed to the application
279in a driver-specific way and can then be used as the mmap offset
280argument.
281
282The GEM core provides a helper method :c:func:`drm_gem_mmap()` to
283handle object mapping. The method can be set directly as the mmap file
284operation handler. It will look up the GEM object based on the offset
285value and set the VMA operations to the :c:type:`struct drm_driver
286<drm_driver>` gem_vm_ops field. Note that
287:c:func:`drm_gem_mmap()` doesn't map memory to userspace, but
288relies on the driver-provided fault handler to map pages individually.
289
290To use :c:func:`drm_gem_mmap()`, drivers must fill the struct
291:c:type:`struct drm_driver <drm_driver>` gem_vm_ops field
292with a pointer to VM operations.
293
Liviu Dudau059c7a52017-01-31 17:41:09 +0000294The VM operations is a :c:type:`struct vm_operations_struct <vm_operations_struct>`
295made up of several fields, the more interesting ones being:
296
297.. code-block:: c
298
299 struct vm_operations_struct {
300 void (*open)(struct vm_area_struct * area);
301 void (*close)(struct vm_area_struct * area);
302 int (*fault)(struct vm_fault *vmf);
303 };
304
Jani Nikula2fa91d12016-06-21 14:49:02 +0300305
306The open and close operations must update the GEM object reference
307count. Drivers can use the :c:func:`drm_gem_vm_open()` and
308:c:func:`drm_gem_vm_close()` helper functions directly as open
309and close handlers.
310
311The fault operation handler is responsible for mapping individual pages
312to userspace when a page fault occurs. Depending on the memory
313allocation scheme, drivers can allocate pages at fault time, or can
314decide to allocate memory for the GEM object at the time the object is
315created.
316
317Drivers that want to map the GEM object upfront instead of handling page
318faults can implement their own mmap file operation handler.
319
Benjamin Gaignard62a0d982017-01-04 10:12:57 +0100320For platforms without MMU the GEM core provides a helper method
321:c:func:`drm_gem_cma_get_unmapped_area`. The mmap() routines will call
322this to get a proposed address for the mapping.
323
324To use :c:func:`drm_gem_cma_get_unmapped_area`, drivers must fill the
325struct :c:type:`struct file_operations <file_operations>` get_unmapped_area
326field with a pointer on :c:func:`drm_gem_cma_get_unmapped_area`.
327
328More detailed information about get_unmapped_area can be found in
329Documentation/nommu-mmap.txt
330
Jani Nikula2fa91d12016-06-21 14:49:02 +0300331Memory Coherency
Daniel Vetter8febdf02016-08-12 22:48:40 +0200332----------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300333
334When mapped to the device or used in a command buffer, backing pages for
335an object are flushed to memory and marked write combined so as to be
336coherent with the GPU. Likewise, if the CPU accesses an object after the
337GPU has finished rendering to the object, then the object must be made
338coherent with the CPU's view of memory, usually involving GPU cache
339flushing of various kinds. This core CPU<->GPU coherency management is
340provided by a device-specific ioctl, which evaluates an object's current
341domain and performs any necessary flushing or synchronization to put the
342object into the desired coherency domain (note that the object may be
343busy, i.e. an active render target; in that case, setting the domain
344blocks the client and waits for rendering to complete before performing
345any necessary flushing operations).
346
347Command Execution
Daniel Vetter8febdf02016-08-12 22:48:40 +0200348-----------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300349
350Perhaps the most important GEM function for GPU devices is providing a
351command execution interface to clients. Client programs construct
352command buffers containing references to previously allocated memory
353objects, and then submit them to GEM. At that point, GEM takes care to
354bind all the objects into the GTT, execute the buffer, and provide
355necessary synchronization between clients accessing the same buffers.
356This often involves evicting some objects from the GTT and re-binding
357others (a fairly expensive operation), and providing relocation support
358which hides fixed GTT offsets from clients. Clients must take care not
359to submit command buffers that reference more objects than can fit in
360the GTT; otherwise, GEM will reject them and no rendering will occur.
361Similarly, if several objects in the buffer require fence registers to
362be allocated for correct rendering (e.g. 2D blits on pre-965 chips),
363care must be taken not to require more fence registers than are
364available to the client. Such resource management should be abstracted
365from the client in libdrm.
366
367GEM Function Reference
368----------------------
369
370.. kernel-doc:: drivers/gpu/drm/drm_gem.c
371 :export:
372
373.. kernel-doc:: include/drm/drm_gem.h
374 :internal:
375
Daniel Vetter8febdf02016-08-12 22:48:40 +0200376GEM CMA Helper Functions Reference
377----------------------------------
378
379.. kernel-doc:: drivers/gpu/drm/drm_gem_cma_helper.c
380 :doc: cma helpers
381
382.. kernel-doc:: drivers/gpu/drm/drm_gem_cma_helper.c
383 :export:
384
385.. kernel-doc:: include/drm/drm_gem_cma_helper.h
386 :internal:
387
Jani Nikula2fa91d12016-06-21 14:49:02 +0300388VMA Offset Manager
Daniel Vetter8febdf02016-08-12 22:48:40 +0200389==================
Jani Nikula2fa91d12016-06-21 14:49:02 +0300390
391.. kernel-doc:: drivers/gpu/drm/drm_vma_manager.c
392 :doc: vma offset manager
393
394.. kernel-doc:: drivers/gpu/drm/drm_vma_manager.c
395 :export:
396
397.. kernel-doc:: include/drm/drm_vma_manager.h
398 :internal:
399
400PRIME Buffer Sharing
Daniel Vetter8febdf02016-08-12 22:48:40 +0200401====================
Jani Nikula2fa91d12016-06-21 14:49:02 +0300402
403PRIME is the cross device buffer sharing framework in drm, originally
404created for the OPTIMUS range of multi-gpu platforms. To userspace PRIME
405buffers are dma-buf based file descriptors.
406
407Overview and Driver Interface
Daniel Vetter8febdf02016-08-12 22:48:40 +0200408-----------------------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300409
410Similar to GEM global names, PRIME file descriptors are also used to
411share buffer objects across processes. They offer additional security:
412as file descriptors must be explicitly sent over UNIX domain sockets to
413be shared between applications, they can't be guessed like the globally
414unique GEM names.
415
416Drivers that support the PRIME API must set the DRIVER_PRIME bit in the
417struct :c:type:`struct drm_driver <drm_driver>`
418driver_features field, and implement the prime_handle_to_fd and
419prime_fd_to_handle operations.
420
421int (\*prime_handle_to_fd)(struct drm_device \*dev, struct drm_file
422\*file_priv, uint32_t handle, uint32_t flags, int \*prime_fd); int
423(\*prime_fd_to_handle)(struct drm_device \*dev, struct drm_file
424\*file_priv, int prime_fd, uint32_t \*handle); Those two operations
425convert a handle to a PRIME file descriptor and vice versa. Drivers must
426use the kernel dma-buf buffer sharing framework to manage the PRIME file
427descriptors. Similar to the mode setting API PRIME is agnostic to the
428underlying buffer object manager, as long as handles are 32bit unsigned
429integers.
430
431While non-GEM drivers must implement the operations themselves, GEM
432drivers must use the :c:func:`drm_gem_prime_handle_to_fd()` and
433:c:func:`drm_gem_prime_fd_to_handle()` helper functions. Those
434helpers rely on the driver gem_prime_export and gem_prime_import
435operations to create a dma-buf instance from a GEM object (dma-buf
436exporter role) and to create a GEM object from a dma-buf instance
437(dma-buf importer role).
438
439struct dma_buf \* (\*gem_prime_export)(struct drm_device \*dev,
440struct drm_gem_object \*obj, int flags); struct drm_gem_object \*
441(\*gem_prime_import)(struct drm_device \*dev, struct dma_buf
442\*dma_buf); These two operations are mandatory for GEM drivers that
443support PRIME.
444
445PRIME Helper Functions
Daniel Vetter8febdf02016-08-12 22:48:40 +0200446----------------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300447
448.. kernel-doc:: drivers/gpu/drm/drm_prime.c
449 :doc: PRIME Helpers
450
451PRIME Function References
452-------------------------
453
454.. kernel-doc:: drivers/gpu/drm/drm_prime.c
455 :export:
456
457DRM MM Range Allocator
Daniel Vetter8febdf02016-08-12 22:48:40 +0200458======================
Jani Nikula2fa91d12016-06-21 14:49:02 +0300459
460Overview
Daniel Vetter8febdf02016-08-12 22:48:40 +0200461--------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300462
463.. kernel-doc:: drivers/gpu/drm/drm_mm.c
464 :doc: Overview
465
466LRU Scan/Eviction Support
Daniel Vetter8febdf02016-08-12 22:48:40 +0200467-------------------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300468
469.. kernel-doc:: drivers/gpu/drm/drm_mm.c
Daniel Vetter05fc0322016-12-29 21:48:23 +0100470 :doc: lru scan roster
Jani Nikula2fa91d12016-06-21 14:49:02 +0300471
472DRM MM Range Allocator Function References
473------------------------------------------
474
475.. kernel-doc:: drivers/gpu/drm/drm_mm.c
476 :export:
477
478.. kernel-doc:: include/drm/drm_mm.h
479 :internal:
Gabriel Krisman Bertazif0e36722017-01-09 19:56:48 -0200480
481DRM Cache Handling
482==================
483
484.. kernel-doc:: drivers/gpu/drm/drm_cache.c
485 :export: