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 .. kernel-doc:: drivers/gpu/drm/ttm/ttm_module.c
34 .. kernel-doc:: include/drm/ttm/ttm_caching.h
37 TTM device object reference
38 ---------------------------
40 .. kernel-doc:: include/drm/ttm/ttm_device.h
43 .. kernel-doc:: drivers/gpu/drm/ttm/ttm_device.c
46 TTM resource placement reference
47 --------------------------------
49 .. kernel-doc:: include/drm/ttm/ttm_placement.h
52 TTM resource object reference
53 -----------------------------
55 .. kernel-doc:: include/drm/ttm/ttm_resource.h
58 .. kernel-doc:: drivers/gpu/drm/ttm/ttm_resource.c
61 TTM TT object reference
62 -----------------------
64 .. kernel-doc:: include/drm/ttm/ttm_tt.h
67 .. kernel-doc:: drivers/gpu/drm/ttm/ttm_tt.c
70 The Graphics Execution Manager (GEM)
71 ====================================
73 The GEM design approach has resulted in a memory manager that doesn't
74 provide full coverage of all (or even all common) use cases in its
75 userspace or kernel API. GEM exposes a set of standard memory-related
76 operations to userspace and a set of helper functions to drivers, and
77 let drivers implement hardware-specific operations with their own
80 The GEM userspace API is described in the `GEM - the Graphics Execution
81 Manager <http://lwn.net/Articles/283798/>`__ article on LWN. While
82 slightly outdated, the document provides a good overview of the GEM API
83 principles. Buffer allocation and read and write operations, described
84 as part of the common GEM API, are currently implemented using
85 driver-specific ioctls.
87 GEM is data-agnostic. It manages abstract buffer objects without knowing
88 what individual buffers contain. APIs that require knowledge of buffer
89 contents or purpose, such as buffer allocation or synchronization
90 primitives, are thus outside of the scope of GEM and must be implemented
91 using driver-specific ioctls.
93 On a fundamental level, GEM involves several operations:
95 - Memory allocation and freeing
97 - Aperture management at command execution time
99 Buffer object allocation is relatively straightforward and largely
100 provided by Linux's shmem layer, which provides memory to back each
103 Device-specific operations, such as command execution, pinning, buffer
104 read & write, mapping, and domain ownership transfers are left to
105 driver-specific ioctls.
110 Drivers that use GEM must set the DRIVER_GEM bit in the struct
111 :c:type:`struct drm_driver <drm_driver>` driver_features
112 field. The DRM core will then automatically initialize the GEM core
113 before calling the load operation. Behind the scene, this will create a
114 DRM Memory Manager object which provides an address space pool for
117 In a KMS configuration, drivers need to allocate and initialize a
118 command ring buffer following core GEM initialization if required by the
119 hardware. UMA devices usually have what is called a "stolen" memory
120 region, which provides space for the initial framebuffer and large,
121 contiguous memory regions required by the device. This space is
122 typically not managed by GEM, and must be initialized separately into
123 its own DRM MM object.
128 GEM splits creation of GEM objects and allocation of the memory that
129 backs them in two distinct operations.
131 GEM objects are represented by an instance of struct :c:type:`struct
132 drm_gem_object <drm_gem_object>`. Drivers usually need to
133 extend GEM objects with private information and thus create a
134 driver-specific GEM object structure type that embeds an instance of
135 struct :c:type:`struct drm_gem_object <drm_gem_object>`.
137 To create a GEM object, a driver allocates memory for an instance of its
138 specific GEM object type and initializes the embedded struct
139 :c:type:`struct drm_gem_object <drm_gem_object>` with a call
140 to drm_gem_object_init(). The function takes a pointer
141 to the DRM device, a pointer to the GEM object and the buffer object
144 GEM uses shmem to allocate anonymous pageable memory.
145 drm_gem_object_init() will create an shmfs file of the
146 requested size and store it into the struct :c:type:`struct
147 drm_gem_object <drm_gem_object>` filp field. The memory is
148 used as either main storage for the object when the graphics hardware
149 uses system memory directly or as a backing store otherwise.
151 Drivers are responsible for the actual physical pages allocation by
152 calling shmem_read_mapping_page_gfp() for each page.
153 Note that they can decide to allocate pages when initializing the GEM
154 object, or to delay allocation until the memory is needed (for instance
155 when a page fault occurs as a result of a userspace memory access or
156 when the driver needs to start a DMA transfer involving the memory).
158 Anonymous pageable memory allocation is not always desired, for instance
159 when the hardware requires physically contiguous system memory as is
160 often the case in embedded devices. Drivers can create GEM objects with
161 no shmfs backing (called private GEM objects) by initializing them with a call
162 to drm_gem_private_object_init() instead of drm_gem_object_init(). Storage for
163 private GEM objects must be managed by drivers.
168 All GEM objects are reference-counted by the GEM core. References can be
169 acquired and release by calling drm_gem_object_get() and drm_gem_object_put()
172 When the last reference to a GEM object is released the GEM core calls
173 the :c:type:`struct drm_gem_object_funcs <gem_object_funcs>` free
174 operation. That operation is mandatory for GEM-enabled drivers and must
175 free the GEM object and all associated resources.
177 void (\*free) (struct drm_gem_object \*obj); Drivers are
178 responsible for freeing all GEM object resources. This includes the
179 resources created by the GEM core, which need to be released with
180 drm_gem_object_release().
185 Communication between userspace and the kernel refers to GEM objects
186 using local handles, global names or, more recently, file descriptors.
187 All of those are 32-bit integer values; the usual Linux kernel limits
188 apply to the file descriptors.
190 GEM handles are local to a DRM file. Applications get a handle to a GEM
191 object through a driver-specific ioctl, and can use that handle to refer
192 to the GEM object in other standard or driver-specific ioctls. Closing a
193 DRM file handle frees all its GEM handles and dereferences the
194 associated GEM objects.
196 To create a handle for a GEM object drivers call drm_gem_handle_create(). The
197 function takes a pointer to the DRM file and the GEM object and returns a
198 locally unique handle. When the handle is no longer needed drivers delete it
199 with a call to drm_gem_handle_delete(). Finally the GEM object associated with a
200 handle can be retrieved by a call to drm_gem_object_lookup().
202 Handles don't take ownership of GEM objects, they only take a reference
203 to the object that will be dropped when the handle is destroyed. To
204 avoid leaking GEM objects, drivers must make sure they drop the
205 reference(s) they own (such as the initial reference taken at object
206 creation time) as appropriate, without any special consideration for the
207 handle. For example, in the particular case of combined GEM object and
208 handle creation in the implementation of the dumb_create operation,
209 drivers must drop the initial reference to the GEM object before
210 returning the handle.
212 GEM names are similar in purpose to handles but are not local to DRM
213 files. They can be passed between processes to reference a GEM object
214 globally. Names can't be used directly to refer to objects in the DRM
215 API, applications must convert handles to names and names to handles
216 using the DRM_IOCTL_GEM_FLINK and DRM_IOCTL_GEM_OPEN ioctls
217 respectively. The conversion is handled by the DRM core without any
218 driver-specific support.
220 GEM also supports buffer sharing with dma-buf file descriptors through
221 PRIME. GEM-based drivers must use the provided helpers functions to
222 implement the exporting and importing correctly. See ?. Since sharing
223 file descriptors is inherently more secure than the easily guessable and
224 global GEM names it is the preferred buffer sharing mechanism. Sharing
225 buffers through GEM names is only supported for legacy userspace.
226 Furthermore PRIME also allows cross-device buffer sharing since it is
232 Because mapping operations are fairly heavyweight GEM favours
233 read/write-like access to buffers, implemented through driver-specific
234 ioctls, over mapping buffers to userspace. However, when random access
235 to the buffer is needed (to perform software rendering for instance),
236 direct access to the object can be more efficient.
238 The mmap system call can't be used directly to map GEM objects, as they
239 don't have their own file handle. Two alternative methods currently
240 co-exist to map GEM objects to userspace. The first method uses a
241 driver-specific ioctl to perform the mapping operation, calling
242 do_mmap() under the hood. This is often considered
243 dubious, seems to be discouraged for new GEM-enabled drivers, and will
244 thus not be described here.
246 The second method uses the mmap system call on the DRM file handle. void
247 \*mmap(void \*addr, size_t length, int prot, int flags, int fd, off_t
248 offset); DRM identifies the GEM object to be mapped by a fake offset
249 passed through the mmap offset argument. Prior to being mapped, a GEM
250 object must thus be associated with a fake offset. To do so, drivers
251 must call drm_gem_create_mmap_offset() on the object.
253 Once allocated, the fake offset value must be passed to the application
254 in a driver-specific way and can then be used as the mmap offset
257 The GEM core provides a helper method drm_gem_mmap() to
258 handle object mapping. The method can be set directly as the mmap file
259 operation handler. It will look up the GEM object based on the offset
260 value and set the VMA operations to the :c:type:`struct drm_driver
261 <drm_driver>` gem_vm_ops field. Note that drm_gem_mmap() doesn't map memory to
262 userspace, but relies on the driver-provided fault handler to map pages
265 To use drm_gem_mmap(), drivers must fill the struct :c:type:`struct drm_driver
266 <drm_driver>` gem_vm_ops field with a pointer to VM operations.
268 The VM operations is a :c:type:`struct vm_operations_struct <vm_operations_struct>`
269 made up of several fields, the more interesting ones being:
273 struct vm_operations_struct {
274 void (*open)(struct vm_area_struct * area);
275 void (*close)(struct vm_area_struct * area);
276 vm_fault_t (*fault)(struct vm_fault *vmf);
280 The open and close operations must update the GEM object reference
281 count. Drivers can use the drm_gem_vm_open() and drm_gem_vm_close() helper
282 functions directly as open and close handlers.
284 The fault operation handler is responsible for mapping individual pages
285 to userspace when a page fault occurs. Depending on the memory
286 allocation scheme, drivers can allocate pages at fault time, or can
287 decide to allocate memory for the GEM object at the time the object is
290 Drivers that want to map the GEM object upfront instead of handling page
291 faults can implement their own mmap file operation handler.
293 For platforms without MMU the GEM core provides a helper method
294 drm_gem_cma_get_unmapped_area(). The mmap() routines will call this to get a
295 proposed address for the mapping.
297 To use drm_gem_cma_get_unmapped_area(), drivers must fill the struct
298 :c:type:`struct file_operations <file_operations>` get_unmapped_area field with
299 a pointer on drm_gem_cma_get_unmapped_area().
301 More detailed information about get_unmapped_area can be found in
302 Documentation/admin-guide/mm/nommu-mmap.rst
307 When mapped to the device or used in a command buffer, backing pages for
308 an object are flushed to memory and marked write combined so as to be
309 coherent with the GPU. Likewise, if the CPU accesses an object after the
310 GPU has finished rendering to the object, then the object must be made
311 coherent with the CPU's view of memory, usually involving GPU cache
312 flushing of various kinds. This core CPU<->GPU coherency management is
313 provided by a device-specific ioctl, which evaluates an object's current
314 domain and performs any necessary flushing or synchronization to put the
315 object into the desired coherency domain (note that the object may be
316 busy, i.e. an active render target; in that case, setting the domain
317 blocks the client and waits for rendering to complete before performing
318 any necessary flushing operations).
323 Perhaps the most important GEM function for GPU devices is providing a
324 command execution interface to clients. Client programs construct
325 command buffers containing references to previously allocated memory
326 objects, and then submit them to GEM. At that point, GEM takes care to
327 bind all the objects into the GTT, execute the buffer, and provide
328 necessary synchronization between clients accessing the same buffers.
329 This often involves evicting some objects from the GTT and re-binding
330 others (a fairly expensive operation), and providing relocation support
331 which hides fixed GTT offsets from clients. Clients must take care not
332 to submit command buffers that reference more objects than can fit in
333 the GTT; otherwise, GEM will reject them and no rendering will occur.
334 Similarly, if several objects in the buffer require fence registers to
335 be allocated for correct rendering (e.g. 2D blits on pre-965 chips),
336 care must be taken not to require more fence registers than are
337 available to the client. Such resource management should be abstracted
338 from the client in libdrm.
340 GEM Function Reference
341 ----------------------
343 .. kernel-doc:: include/drm/drm_gem.h
346 .. kernel-doc:: drivers/gpu/drm/drm_gem.c
349 GEM CMA Helper Functions Reference
350 ----------------------------------
352 .. kernel-doc:: drivers/gpu/drm/drm_gem_cma_helper.c
355 .. kernel-doc:: include/drm/drm_gem_cma_helper.h
358 .. kernel-doc:: drivers/gpu/drm/drm_gem_cma_helper.c
361 GEM SHMEM Helper Function Reference
362 -----------------------------------
364 .. kernel-doc:: drivers/gpu/drm/drm_gem_shmem_helper.c
367 .. kernel-doc:: include/drm/drm_gem_shmem_helper.h
370 .. kernel-doc:: drivers/gpu/drm/drm_gem_shmem_helper.c
373 GEM VRAM Helper Functions Reference
374 -----------------------------------
376 .. kernel-doc:: drivers/gpu/drm/drm_gem_vram_helper.c
379 .. kernel-doc:: include/drm/drm_gem_vram_helper.h
382 .. kernel-doc:: drivers/gpu/drm/drm_gem_vram_helper.c
385 GEM TTM Helper Functions Reference
386 -----------------------------------
388 .. kernel-doc:: drivers/gpu/drm/drm_gem_ttm_helper.c
391 .. kernel-doc:: drivers/gpu/drm/drm_gem_ttm_helper.c
397 .. kernel-doc:: drivers/gpu/drm/drm_vma_manager.c
398 :doc: vma offset manager
400 .. kernel-doc:: include/drm/drm_vma_manager.h
403 .. kernel-doc:: drivers/gpu/drm/drm_vma_manager.c
406 .. _prime_buffer_sharing:
411 PRIME is the cross device buffer sharing framework in drm, originally
412 created for the OPTIMUS range of multi-gpu platforms. To userspace PRIME
413 buffers are dma-buf based file descriptors.
415 Overview and Lifetime Rules
416 ---------------------------
418 .. kernel-doc:: drivers/gpu/drm/drm_prime.c
419 :doc: overview and lifetime rules
421 PRIME Helper Functions
422 ----------------------
424 .. kernel-doc:: drivers/gpu/drm/drm_prime.c
427 PRIME Function References
428 -------------------------
430 .. kernel-doc:: include/drm/drm_prime.h
433 .. kernel-doc:: drivers/gpu/drm/drm_prime.c
436 DRM MM Range Allocator
437 ======================
442 .. kernel-doc:: drivers/gpu/drm/drm_mm.c
445 LRU Scan/Eviction Support
446 -------------------------
448 .. kernel-doc:: drivers/gpu/drm/drm_mm.c
449 :doc: lru scan roster
451 DRM MM Range Allocator Function References
452 ------------------------------------------
454 .. kernel-doc:: include/drm/drm_mm.h
457 .. kernel-doc:: drivers/gpu/drm/drm_mm.c
460 DRM Cache Handling and Fast WC memcpy()
461 =======================================
463 .. kernel-doc:: drivers/gpu/drm/drm_cache.c
467 ===========================
469 .. kernel-doc:: drivers/gpu/drm/drm_syncobj.c
472 .. kernel-doc:: include/drm/drm_syncobj.h
475 .. kernel-doc:: drivers/gpu/drm/drm_syncobj.c
484 .. kernel-doc:: drivers/gpu/drm/scheduler/sched_main.c
487 Scheduler Function References
488 -----------------------------
490 .. kernel-doc:: include/drm/gpu_scheduler.h
493 .. kernel-doc:: drivers/gpu/drm/scheduler/sched_main.c
496 .. kernel-doc:: drivers/gpu/drm/scheduler/sched_entity.c