5 Modern Linux systems require large amount of graphics memory to store
6 frame buffers, textures, vertices and other graphics-related data. Given
7 the very dynamic nature of many of that data, managing graphics memory
8 efficiently is thus crucial for the graphics stack and plays a central
9 role in the DRM infrastructure.
11 The DRM core includes two memory managers, namely Translation Table Maps
12 (TTM) and Graphics Execution Manager (GEM). TTM was the first DRM memory
13 manager to be developed and tried to be a one-size-fits-them all
14 solution. It provides a single userspace API to accommodate the need of
15 all hardware, supporting both Unified Memory Architecture (UMA) devices
16 and devices with dedicated video RAM (i.e. most discrete video cards).
17 This resulted in a large, complex piece of code that turned out to be
18 hard to use for driver development.
20 GEM started as an Intel-sponsored project in reaction to TTM's
21 complexity. Its design philosophy is completely different: instead of
22 providing a solution to every graphics memory-related problems, GEM
23 identified common code between drivers and created a support library to
24 share it. GEM has simpler initialization and execution requirements than
25 TTM, but has no video RAM management capabilities and is thus limited to
28 The Translation Table Manager (TTM)
29 ===================================
31 TTM design background and information belongs here.
37 This section is outdated.
39 Drivers 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
41 initialized global reference to the memory manager. The ttm_bo_driver
42 structure contains several fields with function pointers for
43 initializing the TTM, allocating and freeing memory, waiting for command
44 completion and fence synchronization, and memory migration.
46 The :c:type:`struct drm_global_reference <drm_global_reference>` is made
51 struct drm_global_reference {
52 enum ttm_global_types global_type;
55 int (*init) (struct drm_global_reference *);
56 void (*release) (struct drm_global_reference *);
60 There should be one global reference structure for your memory manager
61 as a whole, and there will be others for each object created by the
62 memory manager at runtime. Your global TTM should have a type of
63 TTM_GLOBAL_TTM_MEM. The size field for the global object should be
64 sizeof(struct ttm_mem_global), and the init and release hooks should
65 point at your driver-specific init and release routines, which probably
66 eventually call ttm_mem_global_init and ttm_mem_global_release,
69 Once your global TTM accounting structure is set up and initialized by
70 calling ttm_global_item_ref() on it, you need to create a buffer
71 object TTM to provide a pool for buffer object allocation by clients and
72 the kernel itself. The type of this object should be
73 TTM_GLOBAL_TTM_BO, and its size should be sizeof(struct
74 ttm_bo_global). Again, driver-specific init and release functions may
75 be provided, likely eventually calling ttm_bo_global_ref_init() and
76 ttm_bo_global_ref_release(), respectively. Also, like the previous
77 object, ttm_global_item_ref() is used to create an initial reference
78 count for the TTM, which will call your initialization function.
80 See the radeon_ttm.c file for an example of usage.
82 The Graphics Execution Manager (GEM)
83 ====================================
85 The GEM design approach has resulted in a memory manager that doesn't
86 provide full coverage of all (or even all common) use cases in its
87 userspace or kernel API. GEM exposes a set of standard memory-related
88 operations to userspace and a set of helper functions to drivers, and
89 let drivers implement hardware-specific operations with their own
92 The GEM userspace API is described in the `GEM - the Graphics Execution
93 Manager <http://lwn.net/Articles/283798/>`__ article on LWN. While
94 slightly outdated, the document provides a good overview of the GEM API
95 principles. Buffer allocation and read and write operations, described
96 as part of the common GEM API, are currently implemented using
97 driver-specific ioctls.
99 GEM is data-agnostic. It manages abstract buffer objects without knowing
100 what individual buffers contain. APIs that require knowledge of buffer
101 contents or purpose, such as buffer allocation or synchronization
102 primitives, are thus outside of the scope of GEM and must be implemented
103 using driver-specific ioctls.
105 On a fundamental level, GEM involves several operations:
107 - Memory allocation and freeing
109 - Aperture management at command execution time
111 Buffer object allocation is relatively straightforward and largely
112 provided by Linux's shmem layer, which provides memory to back each
115 Device-specific operations, such as command execution, pinning, buffer
116 read & write, mapping, and domain ownership transfers are left to
117 driver-specific ioctls.
122 Drivers that use GEM must set the DRIVER_GEM bit in the struct
123 :c:type:`struct drm_driver <drm_driver>` driver_features
124 field. The DRM core will then automatically initialize the GEM core
125 before calling the load operation. Behind the scene, this will create a
126 DRM Memory Manager object which provides an address space pool for
129 In a KMS configuration, drivers need to allocate and initialize a
130 command ring buffer following core GEM initialization if required by the
131 hardware. UMA devices usually have what is called a "stolen" memory
132 region, which provides space for the initial framebuffer and large,
133 contiguous memory regions required by the device. This space is
134 typically not managed by GEM, and must be initialized separately into
135 its own DRM MM object.
140 GEM splits creation of GEM objects and allocation of the memory that
141 backs them in two distinct operations.
143 GEM objects are represented by an instance of struct :c:type:`struct
144 drm_gem_object <drm_gem_object>`. Drivers usually need to
145 extend GEM objects with private information and thus create a
146 driver-specific GEM object structure type that embeds an instance of
147 struct :c:type:`struct drm_gem_object <drm_gem_object>`.
149 To create a GEM object, a driver allocates memory for an instance of its
150 specific GEM object type and initializes the embedded struct
151 :c:type:`struct drm_gem_object <drm_gem_object>` with a call
152 to drm_gem_object_init(). The function takes a pointer
153 to the DRM device, a pointer to the GEM object and the buffer object
156 GEM uses shmem to allocate anonymous pageable memory.
157 drm_gem_object_init() will create an shmfs file of the
158 requested size and store it into the struct :c:type:`struct
159 drm_gem_object <drm_gem_object>` filp field. The memory is
160 used as either main storage for the object when the graphics hardware
161 uses system memory directly or as a backing store otherwise.
163 Drivers are responsible for the actual physical pages allocation by
164 calling shmem_read_mapping_page_gfp() for each page.
165 Note that they can decide to allocate pages when initializing the GEM
166 object, or to delay allocation until the memory is needed (for instance
167 when a page fault occurs as a result of a userspace memory access or
168 when the driver needs to start a DMA transfer involving the memory).
170 Anonymous pageable memory allocation is not always desired, for instance
171 when the hardware requires physically contiguous system memory as is
172 often the case in embedded devices. Drivers can create GEM objects with
173 no shmfs backing (called private GEM objects) by initializing them with a call
174 to drm_gem_private_object_init() instead of drm_gem_object_init(). Storage for
175 private GEM objects must be managed by drivers.
180 All GEM objects are reference-counted by the GEM core. References can be
181 acquired and release by calling drm_gem_object_get() and drm_gem_object_put()
184 When the last reference to a GEM object is released the GEM core calls
185 the :c:type:`struct drm_gem_object_funcs <gem_object_funcs>` free
186 operation. That operation is mandatory for GEM-enabled drivers and must
187 free the GEM object and all associated resources.
189 void (\*free) (struct drm_gem_object \*obj); Drivers are
190 responsible for freeing all GEM object resources. This includes the
191 resources created by the GEM core, which need to be released with
192 drm_gem_object_release().
197 Communication between userspace and the kernel refers to GEM objects
198 using local handles, global names or, more recently, file descriptors.
199 All of those are 32-bit integer values; the usual Linux kernel limits
200 apply to the file descriptors.
202 GEM handles are local to a DRM file. Applications get a handle to a GEM
203 object through a driver-specific ioctl, and can use that handle to refer
204 to the GEM object in other standard or driver-specific ioctls. Closing a
205 DRM file handle frees all its GEM handles and dereferences the
206 associated GEM objects.
208 To create a handle for a GEM object drivers call drm_gem_handle_create(). The
209 function takes a pointer to the DRM file and the GEM object and returns a
210 locally unique handle. When the handle is no longer needed drivers delete it
211 with a call to drm_gem_handle_delete(). Finally the GEM object associated with a
212 handle can be retrieved by a call to drm_gem_object_lookup().
214 Handles don't take ownership of GEM objects, they only take a reference
215 to the object that will be dropped when the handle is destroyed. To
216 avoid leaking GEM objects, drivers must make sure they drop the
217 reference(s) they own (such as the initial reference taken at object
218 creation time) as appropriate, without any special consideration for the
219 handle. For example, in the particular case of combined GEM object and
220 handle creation in the implementation of the dumb_create operation,
221 drivers must drop the initial reference to the GEM object before
222 returning the handle.
224 GEM names are similar in purpose to handles but are not local to DRM
225 files. They can be passed between processes to reference a GEM object
226 globally. Names can't be used directly to refer to objects in the DRM
227 API, applications must convert handles to names and names to handles
228 using the DRM_IOCTL_GEM_FLINK and DRM_IOCTL_GEM_OPEN ioctls
229 respectively. The conversion is handled by the DRM core without any
230 driver-specific support.
232 GEM also supports buffer sharing with dma-buf file descriptors through
233 PRIME. GEM-based drivers must use the provided helpers functions to
234 implement the exporting and importing correctly. See ?. Since sharing
235 file descriptors is inherently more secure than the easily guessable and
236 global GEM names it is the preferred buffer sharing mechanism. Sharing
237 buffers through GEM names is only supported for legacy userspace.
238 Furthermore PRIME also allows cross-device buffer sharing since it is
244 Because mapping operations are fairly heavyweight GEM favours
245 read/write-like access to buffers, implemented through driver-specific
246 ioctls, over mapping buffers to userspace. However, when random access
247 to the buffer is needed (to perform software rendering for instance),
248 direct access to the object can be more efficient.
250 The mmap system call can't be used directly to map GEM objects, as they
251 don't have their own file handle. Two alternative methods currently
252 co-exist to map GEM objects to userspace. The first method uses a
253 driver-specific ioctl to perform the mapping operation, calling
254 do_mmap() under the hood. This is often considered
255 dubious, seems to be discouraged for new GEM-enabled drivers, and will
256 thus not be described here.
258 The second method uses the mmap system call on the DRM file handle. void
259 \*mmap(void \*addr, size_t length, int prot, int flags, int fd, off_t
260 offset); DRM identifies the GEM object to be mapped by a fake offset
261 passed through the mmap offset argument. Prior to being mapped, a GEM
262 object must thus be associated with a fake offset. To do so, drivers
263 must call drm_gem_create_mmap_offset() on the object.
265 Once allocated, the fake offset value must be passed to the application
266 in a driver-specific way and can then be used as the mmap offset
269 The GEM core provides a helper method drm_gem_mmap() to
270 handle object mapping. The method can be set directly as the mmap file
271 operation handler. It will look up the GEM object based on the offset
272 value and set the VMA operations to the :c:type:`struct drm_driver
273 <drm_driver>` gem_vm_ops field. Note that drm_gem_mmap() doesn't map memory to
274 userspace, but relies on the driver-provided fault handler to map pages
277 To use drm_gem_mmap(), drivers must fill the struct :c:type:`struct drm_driver
278 <drm_driver>` gem_vm_ops field with a pointer to VM operations.
280 The VM operations is a :c:type:`struct vm_operations_struct <vm_operations_struct>`
281 made up of several fields, the more interesting ones being:
285 struct vm_operations_struct {
286 void (*open)(struct vm_area_struct * area);
287 void (*close)(struct vm_area_struct * area);
288 vm_fault_t (*fault)(struct vm_fault *vmf);
292 The open and close operations must update the GEM object reference
293 count. Drivers can use the drm_gem_vm_open() and drm_gem_vm_close() helper
294 functions directly as open and close handlers.
296 The fault operation handler is responsible for mapping individual pages
297 to userspace when a page fault occurs. Depending on the memory
298 allocation scheme, drivers can allocate pages at fault time, or can
299 decide to allocate memory for the GEM object at the time the object is
302 Drivers that want to map the GEM object upfront instead of handling page
303 faults can implement their own mmap file operation handler.
305 For platforms without MMU the GEM core provides a helper method
306 drm_gem_cma_get_unmapped_area(). The mmap() routines will call this to get a
307 proposed address for the mapping.
309 To use drm_gem_cma_get_unmapped_area(), drivers must fill the struct
310 :c:type:`struct file_operations <file_operations>` get_unmapped_area field with
311 a pointer on drm_gem_cma_get_unmapped_area().
313 More detailed information about get_unmapped_area can be found in
314 Documentation/admin-guide/mm/nommu-mmap.rst
319 When mapped to the device or used in a command buffer, backing pages for
320 an object are flushed to memory and marked write combined so as to be
321 coherent with the GPU. Likewise, if the CPU accesses an object after the
322 GPU has finished rendering to the object, then the object must be made
323 coherent with the CPU's view of memory, usually involving GPU cache
324 flushing of various kinds. This core CPU<->GPU coherency management is
325 provided by a device-specific ioctl, which evaluates an object's current
326 domain and performs any necessary flushing or synchronization to put the
327 object into the desired coherency domain (note that the object may be
328 busy, i.e. an active render target; in that case, setting the domain
329 blocks the client and waits for rendering to complete before performing
330 any necessary flushing operations).
335 Perhaps the most important GEM function for GPU devices is providing a
336 command execution interface to clients. Client programs construct
337 command buffers containing references to previously allocated memory
338 objects, and then submit them to GEM. At that point, GEM takes care to
339 bind all the objects into the GTT, execute the buffer, and provide
340 necessary synchronization between clients accessing the same buffers.
341 This often involves evicting some objects from the GTT and re-binding
342 others (a fairly expensive operation), and providing relocation support
343 which hides fixed GTT offsets from clients. Clients must take care not
344 to submit command buffers that reference more objects than can fit in
345 the GTT; otherwise, GEM will reject them and no rendering will occur.
346 Similarly, if several objects in the buffer require fence registers to
347 be allocated for correct rendering (e.g. 2D blits on pre-965 chips),
348 care must be taken not to require more fence registers than are
349 available to the client. Such resource management should be abstracted
350 from the client in libdrm.
352 GEM Function Reference
353 ----------------------
355 .. kernel-doc:: include/drm/drm_gem.h
358 .. kernel-doc:: drivers/gpu/drm/drm_gem.c
361 GEM CMA Helper Functions Reference
362 ----------------------------------
364 .. kernel-doc:: drivers/gpu/drm/drm_gem_cma_helper.c
367 .. kernel-doc:: include/drm/drm_gem_cma_helper.h
370 .. kernel-doc:: drivers/gpu/drm/drm_gem_cma_helper.c
373 GEM SHMEM Helper Function Reference
374 -----------------------------------
376 .. kernel-doc:: drivers/gpu/drm/drm_gem_shmem_helper.c
379 .. kernel-doc:: include/drm/drm_gem_shmem_helper.h
382 .. kernel-doc:: drivers/gpu/drm/drm_gem_shmem_helper.c
385 GEM VRAM Helper Functions Reference
386 -----------------------------------
388 .. kernel-doc:: drivers/gpu/drm/drm_gem_vram_helper.c
391 .. kernel-doc:: include/drm/drm_gem_vram_helper.h
394 .. kernel-doc:: drivers/gpu/drm/drm_gem_vram_helper.c
397 GEM TTM Helper Functions Reference
398 -----------------------------------
400 .. kernel-doc:: drivers/gpu/drm/drm_gem_ttm_helper.c
403 .. kernel-doc:: drivers/gpu/drm/drm_gem_ttm_helper.c
409 .. kernel-doc:: drivers/gpu/drm/drm_vma_manager.c
410 :doc: vma offset manager
412 .. kernel-doc:: include/drm/drm_vma_manager.h
415 .. kernel-doc:: drivers/gpu/drm/drm_vma_manager.c
418 .. _prime_buffer_sharing:
423 PRIME is the cross device buffer sharing framework in drm, originally
424 created for the OPTIMUS range of multi-gpu platforms. To userspace PRIME
425 buffers are dma-buf based file descriptors.
427 Overview and Lifetime Rules
428 ---------------------------
430 .. kernel-doc:: drivers/gpu/drm/drm_prime.c
431 :doc: overview and lifetime rules
433 PRIME Helper Functions
434 ----------------------
436 .. kernel-doc:: drivers/gpu/drm/drm_prime.c
439 PRIME Function References
440 -------------------------
442 .. kernel-doc:: include/drm/drm_prime.h
445 .. kernel-doc:: drivers/gpu/drm/drm_prime.c
448 DRM MM Range Allocator
449 ======================
454 .. kernel-doc:: drivers/gpu/drm/drm_mm.c
457 LRU Scan/Eviction Support
458 -------------------------
460 .. kernel-doc:: drivers/gpu/drm/drm_mm.c
461 :doc: lru scan roster
463 DRM MM Range Allocator Function References
464 ------------------------------------------
466 .. kernel-doc:: include/drm/drm_mm.h
469 .. kernel-doc:: drivers/gpu/drm/drm_mm.c
472 DRM Cache Handling and Fast WC memcpy()
473 =======================================
475 .. kernel-doc:: drivers/gpu/drm/drm_cache.c
479 ===========================
481 .. kernel-doc:: drivers/gpu/drm/drm_syncobj.c
484 .. kernel-doc:: include/drm/drm_syncobj.h
487 .. kernel-doc:: drivers/gpu/drm/drm_syncobj.c
496 .. kernel-doc:: drivers/gpu/drm/scheduler/sched_main.c
499 Scheduler Function References
500 -----------------------------
502 .. kernel-doc:: include/drm/gpu_scheduler.h
505 .. kernel-doc:: drivers/gpu/drm/scheduler/sched_main.c