1.. SPDX-License-Identifier: (GPL-2.0+ OR MIT)
  2
  3====================
  4Asynchronous VM_BIND
  5====================
  6
  7Nomenclature:
  8=============
  9
 10* ``VRAM``: On-device memory. Sometimes referred to as device local memory.
 11
 12* ``gpu_vm``: A virtual GPU address space. Typically per process, but
 13  can be shared by multiple processes.
 14
 15* ``VM_BIND``: An operation or a list of operations to modify a gpu_vm using
 16  an IOCTL. The operations include mapping and unmapping system- or
 17  VRAM memory.
 18
 19* ``syncobj``: A container that abstracts synchronization objects. The
 20  synchronization objects can be either generic, like dma-fences or
 21  driver specific. A syncobj typically indicates the type of the
 22  underlying synchronization object.
 23
 24* ``in-syncobj``: Argument to a VM_BIND IOCTL, the VM_BIND operation waits
 25  for these before starting.
 26
 27* ``out-syncobj``: Argument to a VM_BIND_IOCTL, the VM_BIND operation
 28  signals these when the bind operation is complete.
 29
 30* ``dma-fence``: A cross-driver synchronization object. A basic
 31  understanding of dma-fences is required to digest this
 32  document. Please refer to the ``DMA Fences`` section of the
 33  :doc:`dma-buf doc </driver-api/dma-buf>`.
 34
 35* ``memory fence``: A synchronization object, different from a dma-fence.
 36  A memory fence uses the value of a specified memory location to determine
 37  signaled status. A memory fence can be awaited and signaled by both
 38  the GPU and CPU. Memory fences are sometimes referred to as
 39  user-fences, userspace-fences or gpu futexes and do not necessarily obey
 40  the dma-fence rule of signaling within a "reasonable amount of time".
 41  The kernel should thus avoid waiting for memory fences with locks held.
 42
 43* ``long-running workload``: A workload that may take more than the
 44  current stipulated dma-fence maximum signal delay to complete and
 45  which therefore needs to set the gpu_vm or the GPU execution context in
 46  a certain mode that disallows completion dma-fences.
 47
 48* ``exec function``: An exec function is a function that revalidates all
 49  affected gpu_vmas, submits a GPU command batch and registers the
 50  dma_fence representing the GPU command's activity with all affected
 51  dma_resvs. For completeness, although not covered by this document,
 52  it's worth mentioning that an exec function may also be the
 53  revalidation worker that is used by some drivers in compute /
 54  long-running mode.
 55
 56* ``bind context``: A context identifier used for the VM_BIND
 57  operation. VM_BIND operations that use the same bind context can be
 58  assumed, where it matters, to complete in order of submission. No such
 59  assumptions can be made for VM_BIND operations using separate bind contexts.
 60
 61* ``UMD``: User-mode driver.
 62
 63* ``KMD``: Kernel-mode driver.
 64
 65
 66Synchronous / Asynchronous VM_BIND operation
 67============================================
 68
 69Synchronous VM_BIND
 70___________________
 71With Synchronous VM_BIND, the VM_BIND operations all complete before the
 72IOCTL returns. A synchronous VM_BIND takes neither in-fences nor
 73out-fences. Synchronous VM_BIND may block and wait for GPU operations;
 74for example swap-in or clearing, or even previous binds.
 75
 76Asynchronous VM_BIND
 77____________________
 78Asynchronous VM_BIND accepts both in-syncobjs and out-syncobjs. While the
 79IOCTL may return immediately, the VM_BIND operations wait for the in-syncobjs
 80before modifying the GPU page-tables, and signal the out-syncobjs when
 81the modification is done in the sense that the next exec function that
 82awaits for the out-syncobjs will see the change. Errors are reported
 83synchronously.
 84In low-memory situations the implementation may block, performing the
 85VM_BIND synchronously, because there might not be enough memory
 86immediately available for preparing the asynchronous operation.
 87
 88If the VM_BIND IOCTL takes a list or an array of operations as an argument,
 89the in-syncobjs needs to signal before the first operation starts to
 90execute, and the out-syncobjs signal after the last operation
 91completes. Operations in the operation list can be assumed, where it
 92matters, to complete in order.
 93
 94Since asynchronous VM_BIND operations may use dma-fences embedded in
 95out-syncobjs and internally in KMD to signal bind completion,  any
 96memory fences given as VM_BIND in-fences need to be awaited
 97synchronously before the VM_BIND ioctl returns, since dma-fences,
 98required to signal in a reasonable amount of time, can never be made
 99to depend on memory fences that don't have such a restriction.
100
101The purpose of an Asynchronous VM_BIND operation is for user-mode
102drivers to be able to pipeline interleaved gpu_vm modifications and
103exec functions. For long-running workloads, such pipelining of a bind
104operation is not allowed and any in-fences need to be awaited
105synchronously. The reason for this is twofold. First, any memory
106fences gated by a long-running workload and used as in-syncobjs for the
107VM_BIND operation will need to be awaited synchronously anyway (see
108above). Second, any dma-fences used as in-syncobjs for VM_BIND
109operations for long-running workloads will not allow for pipelining
110anyway since long-running workloads don't allow for dma-fences as
111out-syncobjs, so while theoretically possible the use of them is
112questionable and should be rejected until there is a valuable use-case.
113Note that this is not a limitation imposed by dma-fence rules, but
114rather a limitation imposed to keep KMD implementation simple. It does
115not affect using dma-fences as dependencies for the long-running
116workload itself, which is allowed by dma-fence rules, but rather for
117the VM_BIND operation only.
118
119An asynchronous VM_BIND operation may take substantial time to
120complete and signal the out_fence. In particular if the operation is
121deeply pipelined behind other VM_BIND operations and workloads
122submitted using exec functions. In that case, UMD might want to avoid a
123subsequent VM_BIND operation to be queued behind the first one if
124there are no explicit dependencies. In order to circumvent such a queue-up, a
125VM_BIND implementation may allow for VM_BIND contexts to be
126created. For each context, VM_BIND operations will be guaranteed to
127complete in the order they were submitted, but that is not the case
128for VM_BIND operations executing on separate VM_BIND contexts. Instead
129KMD will attempt to execute such VM_BIND operations in parallel but
130leaving no guarantee that they will actually be executed in
131parallel. There may be internal implicit dependencies that only KMD knows
132about, for example page-table structure changes. A way to attempt
133to avoid such internal dependencies is to have different VM_BIND
134contexts use separate regions of a VM.
135
136Also for VM_BINDS for long-running gpu_vms the user-mode driver should typically
137select memory fences as out-fences since that gives greater flexibility for
138the kernel mode driver to inject other operations into the bind /
139unbind operations. Like for example inserting breakpoints into batch
140buffers. The workload execution can then easily be pipelined behind
141the bind completion using the memory out-fence as the signal condition
142for a GPU semaphore embedded by UMD in the workload.
143
144There is no difference in the operations supported or in
145multi-operation support between asynchronous VM_BIND and synchronous VM_BIND.
146
147Multi-operation VM_BIND IOCTL error handling and interrupts
148===========================================================
149
150The VM_BIND operations of the IOCTL may error for various reasons, for
151example due to lack of resources to complete and due to interrupted
152waits.
153In these situations UMD should preferably restart the IOCTL after
154taking suitable action.
155If UMD has over-committed a memory resource, an -ENOSPC error will be
156returned, and UMD may then unbind resources that are not used at the
157moment and rerun the IOCTL. On -EINTR, UMD should simply rerun the
158IOCTL and on -ENOMEM user-space may either attempt to free known
159system memory resources or fail. In case of UMD deciding to fail a
160bind operation, due to an error return, no additional action is needed
161to clean up the failed operation, and the VM is left in the same state
162as it was before the failing IOCTL.
163Unbind operations are guaranteed not to return any errors due to
164resource constraints, but may return errors due to, for example,
165invalid arguments or the gpu_vm being banned.
166In the case an unexpected error happens during the asynchronous bind
167process, the gpu_vm will be banned, and attempts to use it after banning
168will return -ENOENT.
169
170Example: The Xe VM_BIND uAPI
171============================
172
173Starting with the VM_BIND operation struct, the IOCTL call can take
174zero, one or many such operations. A zero number means only the
175synchronization part of the IOCTL is carried out: an asynchronous
176VM_BIND updates the syncobjects, whereas a sync VM_BIND waits for the
177implicit dependencies to be fulfilled.
178
179.. code-block:: c
180
181   struct drm_xe_vm_bind_op {
182	/**
183	 * @obj: GEM object to operate on, MBZ for MAP_USERPTR, MBZ for UNMAP
184	 */
185	__u32 obj;
186
187	/** @pad: MBZ */
188	__u32 pad;
189
190	union {
191		/**
192		 * @obj_offset: Offset into the object for MAP.
193		 */
194		__u64 obj_offset;
195
196		/** @userptr: user virtual address for MAP_USERPTR */
197		__u64 userptr;
198	};
199
200	/**
201	 * @range: Number of bytes from the object to bind to addr, MBZ for UNMAP_ALL
202	 */
203	__u64 range;
204
205	/** @addr: Address to operate on, MBZ for UNMAP_ALL */
206	__u64 addr;
207
208	/**
209	 * @tile_mask: Mask for which tiles to create binds for, 0 == All tiles,
210	 * only applies to creating new VMAs
211	 */
212	__u64 tile_mask;
213
214       /* Map (parts of) an object into the GPU virtual address range.
215    #define XE_VM_BIND_OP_MAP		0x0
216        /* Unmap a GPU virtual address range */
217    #define XE_VM_BIND_OP_UNMAP		0x1
218        /*
219	 * Map a CPU virtual address range into a GPU virtual
220	 * address range.
221	 */
222    #define XE_VM_BIND_OP_MAP_USERPTR	0x2
223        /* Unmap a gem object from the VM. */
224    #define XE_VM_BIND_OP_UNMAP_ALL	0x3
225        /*
226	 * Make the backing memory of an address range resident if
227	 * possible. Note that this doesn't pin backing memory.
228	 */
229    #define XE_VM_BIND_OP_PREFETCH	0x4
230
231        /* Make the GPU map readonly. */
232    #define XE_VM_BIND_FLAG_READONLY	(0x1 << 16)
233	/*
234	 * Valid on a faulting VM only, do the MAP operation immediately rather
235	 * than deferring the MAP to the page fault handler.
236	 */
237    #define XE_VM_BIND_FLAG_IMMEDIATE	(0x1 << 17)
238	/*
239	 * When the NULL flag is set, the page tables are setup with a special
240	 * bit which indicates writes are dropped and all reads return zero.  In
241	 * the future, the NULL flags will only be valid for XE_VM_BIND_OP_MAP
242	 * operations, the BO handle MBZ, and the BO offset MBZ. This flag is
243	 * intended to implement VK sparse bindings.
244	 */
245    #define XE_VM_BIND_FLAG_NULL	(0x1 << 18)
246	/** @op: Operation to perform (lower 16 bits) and flags (upper 16 bits) */
247	__u32 op;
248
249	/** @mem_region: Memory region to prefetch VMA to, instance not a mask */
250	__u32 region;
251
252	/** @reserved: Reserved */
253	__u64 reserved[2];
254   };
255
256
257The VM_BIND IOCTL argument itself, looks like follows. Note that for
258synchronous VM_BIND, the num_syncs and syncs fields must be zero. Here
259the ``exec_queue_id`` field is the VM_BIND context discussed previously
260that is used to facilitate out-of-order VM_BINDs.
261
262.. code-block:: c
263
264    struct drm_xe_vm_bind {
265	/** @extensions: Pointer to the first extension struct, if any */
266	__u64 extensions;
267
268	/** @vm_id: The ID of the VM to bind to */
269	__u32 vm_id;
270
271	/**
272	 * @exec_queue_id: exec_queue_id, must be of class DRM_XE_ENGINE_CLASS_VM_BIND
273	 * and exec queue must have same vm_id. If zero, the default VM bind engine
274	 * is used.
275	 */
276	__u32 exec_queue_id;
277
278	/** @num_binds: number of binds in this IOCTL */
279	__u32 num_binds;
280
281        /* If set, perform an async VM_BIND, if clear a sync VM_BIND */
282    #define XE_VM_BIND_IOCTL_FLAG_ASYNC	(0x1 << 0)
283
284	/** @flag: Flags controlling all operations in this ioctl. */
285	__u32 flags;
286
287	union {
288		/** @bind: used if num_binds == 1 */
289		struct drm_xe_vm_bind_op bind;
290
291		/**
292		 * @vector_of_binds: userptr to array of struct
293		 * drm_xe_vm_bind_op if num_binds > 1
294		 */
295		__u64 vector_of_binds;
296	};
297
298	/** @num_syncs: amount of syncs to wait for or to signal on completion. */
299	__u32 num_syncs;
300
301	/** @pad2: MBZ */
302	__u32 pad2;
303
304	/** @syncs: pointer to struct drm_xe_sync array */
305	__u64 syncs;
306
307	/** @reserved: Reserved */
308	__u64 reserved[2];
309    };