310 lines · plain
1.. SPDX-License-Identifier: (GPL-2.0+ OR MIT)2 3====================4Asynchronous VM_BIND5====================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, but13 can be shared by multiple processes.14 15* ``VM_BIND``: An operation or a list of operations to modify a gpu_vm using16 an IOCTL. The operations include mapping and unmapping system- or17 VRAM memory.18 19* ``syncobj``: A container that abstracts synchronization objects. The20 synchronization objects can be either generic, like dma-fences or21 driver specific. A syncobj typically indicates the type of the22 underlying synchronization object.23 24* ``in-syncobj``: Argument to a VM_BIND IOCTL, the VM_BIND operation waits25 for these before starting.26 27* ``out-syncobj``: Argument to a VM_BIND_IOCTL, the VM_BIND operation28 signals these when the bind operation is complete.29 30* ``dma-fence``: A cross-driver synchronization object. A basic31 understanding of dma-fences is required to digest this32 document. Please refer to the ``DMA Fences`` section of the33 :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 determine37 signaled status. A memory fence can be awaited and signaled by both38 the GPU and CPU. Memory fences are sometimes referred to as39 user-fences, userspace-fences or gpu futexes and do not necessarily obey40 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 the44 current stipulated dma-fence maximum signal delay to complete and45 which therefore needs to set the gpu_vm or the GPU execution context in46 a certain mode that disallows completion dma-fences.47 48* ``exec function``: An exec function is a function that revalidates all49 affected gpu_vmas, submits a GPU command batch and registers the50 dma_fence representing the GPU command's activity with all affected51 dma_resvs. For completeness, although not covered by this document,52 it's worth mentioning that an exec function may also be the53 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_BIND57 operation. VM_BIND operations that use the same bind context can be58 assumed, where it matters, to complete in order of submission. No such59 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 operation67============================================68 69Synchronous VM_BIND70___________________71With Synchronous VM_BIND, the VM_BIND operations all complete before the72IOCTL returns. A synchronous VM_BIND takes neither in-fences nor73out-fences. Synchronous VM_BIND may block and wait for GPU operations;74for example swap-in or clearing, or even previous binds.75 76Asynchronous VM_BIND77____________________78Asynchronous VM_BIND accepts both in-syncobjs and out-syncobjs. While the79IOCTL may return immediately, the VM_BIND operations wait for the in-syncobjs80before modifying the GPU page-tables, and signal the out-syncobjs when81the modification is done in the sense that the next exec function that82awaits for the out-syncobjs will see the change. Errors are reported83synchronously.84In low-memory situations the implementation may block, performing the85VM_BIND synchronously, because there might not be enough memory86immediately 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 to90execute, and the out-syncobjs signal after the last operation91completes. Operations in the operation list can be assumed, where it92matters, to complete in order.93 94Since asynchronous VM_BIND operations may use dma-fences embedded in95out-syncobjs and internally in KMD to signal bind completion, any96memory fences given as VM_BIND in-fences need to be awaited97synchronously before the VM_BIND ioctl returns, since dma-fences,98required to signal in a reasonable amount of time, can never be made99to depend on memory fences that don't have such a restriction.100 101The purpose of an Asynchronous VM_BIND operation is for user-mode102drivers to be able to pipeline interleaved gpu_vm modifications and103exec functions. For long-running workloads, such pipelining of a bind104operation is not allowed and any in-fences need to be awaited105synchronously. The reason for this is twofold. First, any memory106fences gated by a long-running workload and used as in-syncobjs for the107VM_BIND operation will need to be awaited synchronously anyway (see108above). Second, any dma-fences used as in-syncobjs for VM_BIND109operations for long-running workloads will not allow for pipelining110anyway since long-running workloads don't allow for dma-fences as111out-syncobjs, so while theoretically possible the use of them is112questionable and should be rejected until there is a valuable use-case.113Note that this is not a limitation imposed by dma-fence rules, but114rather a limitation imposed to keep KMD implementation simple. It does115not affect using dma-fences as dependencies for the long-running116workload itself, which is allowed by dma-fence rules, but rather for117the VM_BIND operation only.118 119An asynchronous VM_BIND operation may take substantial time to120complete and signal the out_fence. In particular if the operation is121deeply pipelined behind other VM_BIND operations and workloads122submitted using exec functions. In that case, UMD might want to avoid a123subsequent VM_BIND operation to be queued behind the first one if124there are no explicit dependencies. In order to circumvent such a queue-up, a125VM_BIND implementation may allow for VM_BIND contexts to be126created. For each context, VM_BIND operations will be guaranteed to127complete in the order they were submitted, but that is not the case128for VM_BIND operations executing on separate VM_BIND contexts. Instead129KMD will attempt to execute such VM_BIND operations in parallel but130leaving no guarantee that they will actually be executed in131parallel. There may be internal implicit dependencies that only KMD knows132about, for example page-table structure changes. A way to attempt133to avoid such internal dependencies is to have different VM_BIND134contexts use separate regions of a VM.135 136Also for VM_BINDS for long-running gpu_vms the user-mode driver should typically137select memory fences as out-fences since that gives greater flexibility for138the kernel mode driver to inject other operations into the bind /139unbind operations. Like for example inserting breakpoints into batch140buffers. The workload execution can then easily be pipelined behind141the bind completion using the memory out-fence as the signal condition142for a GPU semaphore embedded by UMD in the workload.143 144There is no difference in the operations supported or in145multi-operation support between asynchronous VM_BIND and synchronous VM_BIND.146 147Multi-operation VM_BIND IOCTL error handling and interrupts148===========================================================149 150The VM_BIND operations of the IOCTL may error for various reasons, for151example due to lack of resources to complete and due to interrupted152waits.153In these situations UMD should preferably restart the IOCTL after154taking suitable action.155If UMD has over-committed a memory resource, an -ENOSPC error will be156returned, and UMD may then unbind resources that are not used at the157moment and rerun the IOCTL. On -EINTR, UMD should simply rerun the158IOCTL and on -ENOMEM user-space may either attempt to free known159system memory resources or fail. In case of UMD deciding to fail a160bind operation, due to an error return, no additional action is needed161to clean up the failed operation, and the VM is left in the same state162as it was before the failing IOCTL.163Unbind operations are guaranteed not to return any errors due to164resource 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 bind167process, the gpu_vm will be banned, and attempts to use it after banning168will return -ENOENT.169 170Example: The Xe VM_BIND uAPI171============================172 173Starting with the VM_BIND operation struct, the IOCTL call can take174zero, one or many such operations. A zero number means only the175synchronization part of the IOCTL is carried out: an asynchronous176VM_BIND updates the syncobjects, whereas a sync VM_BIND waits for the177implicit dependencies to be fulfilled.178 179.. code-block:: c180 181 struct drm_xe_vm_bind_op {182 /**183 * @obj: GEM object to operate on, MBZ for MAP_USERPTR, MBZ for UNMAP184 */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_ALL202 */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 VMAs211 */212 __u64 tile_mask;213 214 /* Map (parts of) an object into the GPU virtual address range.215 #define XE_VM_BIND_OP_MAP 0x0216 /* Unmap a GPU virtual address range */217 #define XE_VM_BIND_OP_UNMAP 0x1218 /*219 * Map a CPU virtual address range into a GPU virtual220 * address range.221 */222 #define XE_VM_BIND_OP_MAP_USERPTR 0x2223 /* Unmap a gem object from the VM. */224 #define XE_VM_BIND_OP_UNMAP_ALL 0x3225 /*226 * Make the backing memory of an address range resident if227 * possible. Note that this doesn't pin backing memory.228 */229 #define XE_VM_BIND_OP_PREFETCH 0x4230 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 rather235 * 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 special240 * bit which indicates writes are dropped and all reads return zero. In241 * the future, the NULL flags will only be valid for XE_VM_BIND_OP_MAP242 * operations, the BO handle MBZ, and the BO offset MBZ. This flag is243 * 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 for258synchronous VM_BIND, the num_syncs and syncs fields must be zero. Here259the ``exec_queue_id`` field is the VM_BIND context discussed previously260that is used to facilitate out-of-order VM_BINDs.261 262.. code-block:: c263 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_BIND273 * and exec queue must have same vm_id. If zero, the default VM bind engine274 * 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 struct293 * drm_xe_vm_bind_op if num_binds > 1294 */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 };310