blob: cb5daffcd6be2167c77c2f1da6dd8feb0ce40050 [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**
37
38 This section is outdated.
39
40Drivers wishing to support TTM must fill out a drm_bo_driver
41structure. The structure contains several fields with function pointers
42for initializing the TTM, allocating and freeing memory, waiting for
43command completion and fence synchronization, and memory migration. See
44the radeon_ttm.c file for an example of usage.
45
46The ttm_global_reference structure is made up of several fields:
47
Jani Nikula29849a62016-11-03 11:44:23 +020048.. code-block:: c
Jani Nikula2fa91d12016-06-21 14:49:02 +030049
50 struct ttm_global_reference {
51 enum ttm_global_types global_type;
52 size_t size;
53 void *object;
54 int (*init) (struct ttm_global_reference *);
55 void (*release) (struct ttm_global_reference *);
56 };
57
58
59There should be one global reference structure for your memory manager
60as a whole, and there will be others for each object created by the
61memory manager at runtime. Your global TTM should have a type of
62TTM_GLOBAL_TTM_MEM. The size field for the global object should be
63sizeof(struct ttm_mem_global), and the init and release hooks should
64point at your driver-specific init and release routines, which probably
65eventually call ttm_mem_global_init and ttm_mem_global_release,
66respectively.
67
68Once your global TTM accounting structure is set up and initialized by
69calling ttm_global_item_ref() on it, you need to create a buffer
70object TTM to provide a pool for buffer object allocation by clients and
71the kernel itself. The type of this object should be
72TTM_GLOBAL_TTM_BO, and its size should be sizeof(struct
73ttm_bo_global). Again, driver-specific init and release functions may
74be provided, likely eventually calling ttm_bo_global_init() and
75ttm_bo_global_release(), respectively. Also, like the previous
76object, ttm_global_item_ref() is used to create an initial reference
77count for the TTM, which will call your initialization function.
78
79The Graphics Execution Manager (GEM)
Daniel Vetter8febdf02016-08-12 22:48:40 +020080====================================
Jani Nikula2fa91d12016-06-21 14:49:02 +030081
82The GEM design approach has resulted in a memory manager that doesn't
83provide full coverage of all (or even all common) use cases in its
84userspace or kernel API. GEM exposes a set of standard memory-related
85operations to userspace and a set of helper functions to drivers, and
86let drivers implement hardware-specific operations with their own
87private API.
88
89The GEM userspace API is described in the `GEM - the Graphics Execution
90Manager <http://lwn.net/Articles/283798/>`__ article on LWN. While
91slightly outdated, the document provides a good overview of the GEM API
92principles. Buffer allocation and read and write operations, described
93as part of the common GEM API, are currently implemented using
94driver-specific ioctls.
95
96GEM is data-agnostic. It manages abstract buffer objects without knowing
97what individual buffers contain. APIs that require knowledge of buffer
98contents or purpose, such as buffer allocation or synchronization
99primitives, are thus outside of the scope of GEM and must be implemented
100using driver-specific ioctls.
101
102On a fundamental level, GEM involves several operations:
103
104- Memory allocation and freeing
105- Command execution
106- Aperture management at command execution time
107
108Buffer object allocation is relatively straightforward and largely
109provided by Linux's shmem layer, which provides memory to back each
110object.
111
112Device-specific operations, such as command execution, pinning, buffer
113read & write, mapping, and domain ownership transfers are left to
114driver-specific ioctls.
115
116GEM Initialization
Daniel Vetter8febdf02016-08-12 22:48:40 +0200117------------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300118
119Drivers that use GEM must set the DRIVER_GEM bit in the struct
120:c:type:`struct drm_driver <drm_driver>` driver_features
121field. The DRM core will then automatically initialize the GEM core
122before calling the load operation. Behind the scene, this will create a
123DRM Memory Manager object which provides an address space pool for
124object allocation.
125
126In a KMS configuration, drivers need to allocate and initialize a
127command ring buffer following core GEM initialization if required by the
128hardware. UMA devices usually have what is called a "stolen" memory
129region, which provides space for the initial framebuffer and large,
130contiguous memory regions required by the device. This space is
131typically not managed by GEM, and must be initialized separately into
132its own DRM MM object.
133
134GEM Objects Creation
Daniel Vetter8febdf02016-08-12 22:48:40 +0200135--------------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300136
137GEM splits creation of GEM objects and allocation of the memory that
138backs them in two distinct operations.
139
140GEM objects are represented by an instance of struct :c:type:`struct
141drm_gem_object <drm_gem_object>`. Drivers usually need to
142extend GEM objects with private information and thus create a
143driver-specific GEM object structure type that embeds an instance of
144struct :c:type:`struct drm_gem_object <drm_gem_object>`.
145
146To create a GEM object, a driver allocates memory for an instance of its
147specific GEM object type and initializes the embedded struct
148:c:type:`struct drm_gem_object <drm_gem_object>` with a call
149to :c:func:`drm_gem_object_init()`. The function takes a pointer
150to the DRM device, a pointer to the GEM object and the buffer object
151size in bytes.
152
153GEM uses shmem to allocate anonymous pageable memory.
154:c:func:`drm_gem_object_init()` will create an shmfs file of the
155requested size and store it into the struct :c:type:`struct
156drm_gem_object <drm_gem_object>` filp field. The memory is
157used as either main storage for the object when the graphics hardware
158uses system memory directly or as a backing store otherwise.
159
160Drivers are responsible for the actual physical pages allocation by
161calling :c:func:`shmem_read_mapping_page_gfp()` for each page.
162Note that they can decide to allocate pages when initializing the GEM
163object, or to delay allocation until the memory is needed (for instance
164when a page fault occurs as a result of a userspace memory access or
165when the driver needs to start a DMA transfer involving the memory).
166
167Anonymous pageable memory allocation is not always desired, for instance
168when the hardware requires physically contiguous system memory as is
169often the case in embedded devices. Drivers can create GEM objects with
170no shmfs backing (called private GEM objects) by initializing them with
171a call to :c:func:`drm_gem_private_object_init()` instead of
172:c:func:`drm_gem_object_init()`. Storage for private GEM objects
173must be managed by drivers.
174
175GEM Objects Lifetime
Daniel Vetter8febdf02016-08-12 22:48:40 +0200176--------------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300177
178All GEM objects are reference-counted by the GEM core. References can be
179acquired and release by :c:func:`calling
180drm_gem_object_reference()` and
181:c:func:`drm_gem_object_unreference()` respectively. The caller
182must hold the :c:type:`struct drm_device <drm_device>`
183struct_mutex lock when calling
184:c:func:`drm_gem_object_reference()`. As a convenience, GEM
185provides :c:func:`drm_gem_object_unreference_unlocked()`
186functions that can be called without holding the lock.
187
188When the last reference to a GEM object is released the GEM core calls
189the :c:type:`struct drm_driver <drm_driver>` gem_free_object
190operation. That operation is mandatory for GEM-enabled drivers and must
191free the GEM object and all associated resources.
192
193void (\*gem_free_object) (struct drm_gem_object \*obj); Drivers are
194responsible for freeing all GEM object resources. This includes the
195resources created by the GEM core, which need to be released with
196:c:func:`drm_gem_object_release()`.
197
198GEM Objects Naming
Daniel Vetter8febdf02016-08-12 22:48:40 +0200199------------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300200
201Communication between userspace and the kernel refers to GEM objects
202using local handles, global names or, more recently, file descriptors.
203All of those are 32-bit integer values; the usual Linux kernel limits
204apply to the file descriptors.
205
206GEM handles are local to a DRM file. Applications get a handle to a GEM
207object through a driver-specific ioctl, and can use that handle to refer
208to the GEM object in other standard or driver-specific ioctls. Closing a
209DRM file handle frees all its GEM handles and dereferences the
210associated GEM objects.
211
212To create a handle for a GEM object drivers call
213:c:func:`drm_gem_handle_create()`. The function takes a pointer
214to the DRM file and the GEM object and returns a locally unique handle.
215When the handle is no longer needed drivers delete it with a call to
216:c:func:`drm_gem_handle_delete()`. Finally the GEM object
217associated with a handle can be retrieved by a call to
218:c:func:`drm_gem_object_lookup()`.
219
220Handles don't take ownership of GEM objects, they only take a reference
221to the object that will be dropped when the handle is destroyed. To
222avoid leaking GEM objects, drivers must make sure they drop the
223reference(s) they own (such as the initial reference taken at object
224creation time) as appropriate, without any special consideration for the
225handle. For example, in the particular case of combined GEM object and
226handle creation in the implementation of the dumb_create operation,
227drivers must drop the initial reference to the GEM object before
228returning the handle.
229
230GEM names are similar in purpose to handles but are not local to DRM
231files. They can be passed between processes to reference a GEM object
232globally. Names can't be used directly to refer to objects in the DRM
233API, applications must convert handles to names and names to handles
234using the DRM_IOCTL_GEM_FLINK and DRM_IOCTL_GEM_OPEN ioctls
235respectively. The conversion is handled by the DRM core without any
236driver-specific support.
237
238GEM also supports buffer sharing with dma-buf file descriptors through
239PRIME. GEM-based drivers must use the provided helpers functions to
240implement the exporting and importing correctly. See ?. Since sharing
241file descriptors is inherently more secure than the easily guessable and
242global GEM names it is the preferred buffer sharing mechanism. Sharing
243buffers through GEM names is only supported for legacy userspace.
244Furthermore PRIME also allows cross-device buffer sharing since it is
245based on dma-bufs.
246
247GEM Objects Mapping
Daniel Vetter8febdf02016-08-12 22:48:40 +0200248-------------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300249
250Because mapping operations are fairly heavyweight GEM favours
251read/write-like access to buffers, implemented through driver-specific
252ioctls, over mapping buffers to userspace. However, when random access
253to the buffer is needed (to perform software rendering for instance),
254direct access to the object can be more efficient.
255
256The mmap system call can't be used directly to map GEM objects, as they
257don't have their own file handle. Two alternative methods currently
258co-exist to map GEM objects to userspace. The first method uses a
259driver-specific ioctl to perform the mapping operation, calling
260:c:func:`do_mmap()` under the hood. This is often considered
261dubious, seems to be discouraged for new GEM-enabled drivers, and will
262thus not be described here.
263
264The second method uses the mmap system call on the DRM file handle. void
265\*mmap(void \*addr, size_t length, int prot, int flags, int fd, off_t
266offset); DRM identifies the GEM object to be mapped by a fake offset
267passed through the mmap offset argument. Prior to being mapped, a GEM
268object must thus be associated with a fake offset. To do so, drivers
269must call :c:func:`drm_gem_create_mmap_offset()` on the object.
270
271Once allocated, the fake offset value must be passed to the application
272in a driver-specific way and can then be used as the mmap offset
273argument.
274
275The GEM core provides a helper method :c:func:`drm_gem_mmap()` to
276handle object mapping. The method can be set directly as the mmap file
277operation handler. It will look up the GEM object based on the offset
278value and set the VMA operations to the :c:type:`struct drm_driver
279<drm_driver>` gem_vm_ops field. Note that
280:c:func:`drm_gem_mmap()` doesn't map memory to userspace, but
281relies on the driver-provided fault handler to map pages individually.
282
283To use :c:func:`drm_gem_mmap()`, drivers must fill the struct
284:c:type:`struct drm_driver <drm_driver>` gem_vm_ops field
285with a pointer to VM operations.
286
287struct vm_operations_struct \*gem_vm_ops struct
288vm_operations_struct { void (\*open)(struct vm_area_struct \* area);
289void (\*close)(struct vm_area_struct \* area); int (\*fault)(struct
290vm_area_struct \*vma, struct vm_fault \*vmf); };
291
292The open and close operations must update the GEM object reference
293count. Drivers can use the :c:func:`drm_gem_vm_open()` and
294:c:func:`drm_gem_vm_close()` helper functions directly as open
295and close handlers.
296
297The fault operation handler is responsible for mapping individual pages
298to userspace when a page fault occurs. Depending on the memory
299allocation scheme, drivers can allocate pages at fault time, or can
300decide to allocate memory for the GEM object at the time the object is
301created.
302
303Drivers that want to map the GEM object upfront instead of handling page
304faults can implement their own mmap file operation handler.
305
306Memory Coherency
Daniel Vetter8febdf02016-08-12 22:48:40 +0200307----------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300308
309When mapped to the device or used in a command buffer, backing pages for
310an object are flushed to memory and marked write combined so as to be
311coherent with the GPU. Likewise, if the CPU accesses an object after the
312GPU has finished rendering to the object, then the object must be made
313coherent with the CPU's view of memory, usually involving GPU cache
314flushing of various kinds. This core CPU<->GPU coherency management is
315provided by a device-specific ioctl, which evaluates an object's current
316domain and performs any necessary flushing or synchronization to put the
317object into the desired coherency domain (note that the object may be
318busy, i.e. an active render target; in that case, setting the domain
319blocks the client and waits for rendering to complete before performing
320any necessary flushing operations).
321
322Command Execution
Daniel Vetter8febdf02016-08-12 22:48:40 +0200323-----------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300324
325Perhaps the most important GEM function for GPU devices is providing a
326command execution interface to clients. Client programs construct
327command buffers containing references to previously allocated memory
328objects, and then submit them to GEM. At that point, GEM takes care to
329bind all the objects into the GTT, execute the buffer, and provide
330necessary synchronization between clients accessing the same buffers.
331This often involves evicting some objects from the GTT and re-binding
332others (a fairly expensive operation), and providing relocation support
333which hides fixed GTT offsets from clients. Clients must take care not
334to submit command buffers that reference more objects than can fit in
335the GTT; otherwise, GEM will reject them and no rendering will occur.
336Similarly, if several objects in the buffer require fence registers to
337be allocated for correct rendering (e.g. 2D blits on pre-965 chips),
338care must be taken not to require more fence registers than are
339available to the client. Such resource management should be abstracted
340from the client in libdrm.
341
342GEM Function Reference
343----------------------
344
345.. kernel-doc:: drivers/gpu/drm/drm_gem.c
346 :export:
347
348.. kernel-doc:: include/drm/drm_gem.h
349 :internal:
350
Daniel Vetter8febdf02016-08-12 22:48:40 +0200351GEM CMA Helper Functions Reference
352----------------------------------
353
354.. kernel-doc:: drivers/gpu/drm/drm_gem_cma_helper.c
355 :doc: cma helpers
356
357.. kernel-doc:: drivers/gpu/drm/drm_gem_cma_helper.c
358 :export:
359
360.. kernel-doc:: include/drm/drm_gem_cma_helper.h
361 :internal:
362
Jani Nikula2fa91d12016-06-21 14:49:02 +0300363VMA Offset Manager
Daniel Vetter8febdf02016-08-12 22:48:40 +0200364==================
Jani Nikula2fa91d12016-06-21 14:49:02 +0300365
366.. kernel-doc:: drivers/gpu/drm/drm_vma_manager.c
367 :doc: vma offset manager
368
369.. kernel-doc:: drivers/gpu/drm/drm_vma_manager.c
370 :export:
371
372.. kernel-doc:: include/drm/drm_vma_manager.h
373 :internal:
374
375PRIME Buffer Sharing
Daniel Vetter8febdf02016-08-12 22:48:40 +0200376====================
Jani Nikula2fa91d12016-06-21 14:49:02 +0300377
378PRIME is the cross device buffer sharing framework in drm, originally
379created for the OPTIMUS range of multi-gpu platforms. To userspace PRIME
380buffers are dma-buf based file descriptors.
381
382Overview and Driver Interface
Daniel Vetter8febdf02016-08-12 22:48:40 +0200383-----------------------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300384
385Similar to GEM global names, PRIME file descriptors are also used to
386share buffer objects across processes. They offer additional security:
387as file descriptors must be explicitly sent over UNIX domain sockets to
388be shared between applications, they can't be guessed like the globally
389unique GEM names.
390
391Drivers that support the PRIME API must set the DRIVER_PRIME bit in the
392struct :c:type:`struct drm_driver <drm_driver>`
393driver_features field, and implement the prime_handle_to_fd and
394prime_fd_to_handle operations.
395
396int (\*prime_handle_to_fd)(struct drm_device \*dev, struct drm_file
397\*file_priv, uint32_t handle, uint32_t flags, int \*prime_fd); int
398(\*prime_fd_to_handle)(struct drm_device \*dev, struct drm_file
399\*file_priv, int prime_fd, uint32_t \*handle); Those two operations
400convert a handle to a PRIME file descriptor and vice versa. Drivers must
401use the kernel dma-buf buffer sharing framework to manage the PRIME file
402descriptors. Similar to the mode setting API PRIME is agnostic to the
403underlying buffer object manager, as long as handles are 32bit unsigned
404integers.
405
406While non-GEM drivers must implement the operations themselves, GEM
407drivers must use the :c:func:`drm_gem_prime_handle_to_fd()` and
408:c:func:`drm_gem_prime_fd_to_handle()` helper functions. Those
409helpers rely on the driver gem_prime_export and gem_prime_import
410operations to create a dma-buf instance from a GEM object (dma-buf
411exporter role) and to create a GEM object from a dma-buf instance
412(dma-buf importer role).
413
414struct dma_buf \* (\*gem_prime_export)(struct drm_device \*dev,
415struct drm_gem_object \*obj, int flags); struct drm_gem_object \*
416(\*gem_prime_import)(struct drm_device \*dev, struct dma_buf
417\*dma_buf); These two operations are mandatory for GEM drivers that
418support PRIME.
419
420PRIME Helper Functions
Daniel Vetter8febdf02016-08-12 22:48:40 +0200421----------------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300422
423.. kernel-doc:: drivers/gpu/drm/drm_prime.c
424 :doc: PRIME Helpers
425
426PRIME Function References
427-------------------------
428
429.. kernel-doc:: drivers/gpu/drm/drm_prime.c
430 :export:
431
432DRM MM Range Allocator
Daniel Vetter8febdf02016-08-12 22:48:40 +0200433======================
Jani Nikula2fa91d12016-06-21 14:49:02 +0300434
435Overview
Daniel Vetter8febdf02016-08-12 22:48:40 +0200436--------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300437
438.. kernel-doc:: drivers/gpu/drm/drm_mm.c
439 :doc: Overview
440
441LRU Scan/Eviction Support
Daniel Vetter8febdf02016-08-12 22:48:40 +0200442-------------------------
Jani Nikula2fa91d12016-06-21 14:49:02 +0300443
444.. kernel-doc:: drivers/gpu/drm/drm_mm.c
445 :doc: lru scan roaster
446
447DRM MM Range Allocator Function References
448------------------------------------------
449
450.. kernel-doc:: drivers/gpu/drm/drm_mm.c
451 :export:
452
453.. kernel-doc:: include/drm/drm_mm.h
454 :internal: