/* SPDX-License-Identifier: MIT */ /* * Copyright © 2022 Intel Corporation */ #ifndef _XE_VM_DOC_H_ #define _XE_VM_DOC_H_ /** * DOC: XE VM (user address space) * * VM creation * =========== * * Allocate a physical page for root of the page table structure, create default * bind engine, and return a handle to the user. * * Scratch page * ------------ * * If the VM is created with the flag, DRM_XE_VM_CREATE_FLAG_SCRATCH_PAGE, set the * entire page table structure defaults pointing to blank page allocated by the * VM. Invalid memory access rather than fault just read / write to this page. * * VM bind (create GPU mapping for a BO or userptr) * ================================================ * * Creates GPU mapings for a BO or userptr within a VM. VM binds uses the same * in / out fence interface (struct drm_xe_sync) as execs which allows users to * think of binds and execs as more or less the same operation. * * Operations * ---------- * * DRM_XE_VM_BIND_OP_MAP - Create mapping for a BO * DRM_XE_VM_BIND_OP_UNMAP - Destroy mapping for a BO / userptr * DRM_XE_VM_BIND_OP_MAP_USERPTR - Create mapping for userptr * * Implementation details * ~~~~~~~~~~~~~~~~~~~~~~ * * All bind operations are implemented via a hybrid approach of using the CPU * and GPU to modify page tables. If a new physical page is allocated in the * page table structure we populate that page via the CPU and insert that new * page into the existing page table structure via a GPU job. Also any existing * pages in the page table structure that need to be modified also are updated * via the GPU job. As the root physical page is prealloced on VM creation our * GPU job will always have at least 1 update. The in / out fences are passed to * this job so again this is conceptually the same as an exec. * * Very simple example of few binds on an empty VM with 48 bits of address space * and the resulting operations: * * .. code-block:: * * bind BO0 0x0-0x1000 * alloc page level 3a, program PTE[0] to BO0 phys address (CPU) * alloc page level 2, program PDE[0] page level 3a phys address (CPU) * alloc page level 1, program PDE[0] page level 2 phys address (CPU) * update root PDE[0] to page level 1 phys address (GPU) * * bind BO1 0x201000-0x202000 * alloc page level 3b, program PTE[1] to BO1 phys address (CPU) * update page level 2 PDE[1] to page level 3b phys address (GPU) * * bind BO2 0x1ff000-0x201000 * update page level 3a PTE[511] to BO2 phys addres (GPU) * update page level 3b PTE[0] to BO2 phys addres + 0x1000 (GPU) * * GPU bypass * ~~~~~~~~~~ * * In the above example the steps using the GPU can be converted to CPU if the * bind can be done immediately (all in-fences satisfied, VM dma-resv kernel * slot is idle). * * Address space * ------------- * * Depending on platform either 48 or 57 bits of address space is supported. * * Page sizes * ---------- * * The minimum page size is either 4k or 64k depending on platform and memory * placement (sysmem vs. VRAM). We enforce that binds must be aligned to the * minimum page size. * * Larger pages (2M or 1GB) can be used for BOs in VRAM, the BO physical address * is aligned to the larger pages size, and VA is aligned to the larger page * size. Larger pages for userptrs / BOs in sysmem should be possible but is not * yet implemented. * * Sync error handling mode * ------------------------ * * In both modes during the bind IOCTL the user input is validated. In sync * error handling mode the newly bound BO is validated (potentially moved back * to a region of memory where is can be used), page tables are updated by the * CPU and the job to do the GPU binds is created in the IOCTL itself. This step * can fail due to memory pressure. The user can recover by freeing memory and * trying this operation again. * * Async error handling mode * ------------------------- * * In async error handling the step of validating the BO, updating page tables, * and generating a job are deferred to an async worker. As this step can now * fail after the IOCTL has reported success we need an error handling flow for * which the user can recover from. * * The solution is for a user to register a user address with the VM which the * VM uses to report errors to. The ufence wait interface can be used to wait on * a VM going into an error state. Once an error is reported the VM's async * worker is paused. While the VM's async worker is paused sync, * DRM_XE_VM_BIND_OP_UNMAP operations are allowed (this can free memory). Once the * uses believe the error state is fixed, the async worker can be resumed via * XE_VM_BIND_OP_RESTART operation. When VM async bind work is restarted, the * first operation processed is the operation that caused the original error. * * Bind queues / engines * --------------------- * * Think of the case where we have two bind operations A + B and are submitted * in that order. A has in fences while B has none. If using a single bind * queue, B is now blocked on A's in fences even though it is ready to run. This * example is a real use case for VK sparse binding. We work around this * limitation by implementing bind engines. * * In the bind IOCTL the user can optionally pass in an engine ID which must map * to an engine which is of the special class DRM_XE_ENGINE_CLASS_VM_BIND. * Underneath this is a really virtual engine that can run on any of the copy * hardware engines. The job(s) created each IOCTL are inserted into this * engine's ring. In the example above if A and B have different bind engines B * is free to pass A. If the engine ID field is omitted, the default bind queue * for the VM is used. * * TODO: Explain race in issue 41 and how we solve it * * Array of bind operations * ------------------------ * * The uAPI allows multiple binds operations to be passed in via a user array, * of struct drm_xe_vm_bind_op, in a single VM bind IOCTL. This interface * matches the VK sparse binding API. The implementation is rather simple, parse * the array into a list of operations, pass the in fences to the first operation, * and pass the out fences to the last operation. The ordered nature of a bind * engine makes this possible. * * Munmap semantics for unbinds * ---------------------------- * * Munmap allows things like: * * .. code-block:: * * 0x0000-0x2000 and 0x3000-0x5000 have mappings * Munmap 0x1000-0x4000, results in mappings 0x0000-0x1000 and 0x4000-0x5000 * * To support this semantic in the above example we decompose the above example * into 4 operations: * * .. code-block:: * * unbind 0x0000-0x2000 * unbind 0x3000-0x5000 * rebind 0x0000-0x1000 * rebind 0x4000-0x5000 * * Why not just do a partial unbind of 0x1000-0x2000 and 0x3000-0x4000? This * falls apart when using large pages at the edges and the unbind forces us to * use a smaller page size. For simplity we always issue a set of unbinds * unmapping anything in the range and at most 2 rebinds on the edges. * * Similar to an array of binds, in fences are passed to the first operation and * out fences are signaled on the last operation. * * In this example there is a window of time where 0x0000-0x1000 and * 0x4000-0x5000 are invalid but the user didn't ask for these addresses to be * removed from the mapping. To work around this we treat any munmap style * unbinds which require a rebind as a kernel operations (BO eviction or userptr * invalidation). The first operation waits on the VM's * DMA_RESV_USAGE_PREEMPT_FENCE slots (waits for all pending jobs on VM to * complete / triggers preempt fences) and the last operation is installed in * the VM's DMA_RESV_USAGE_KERNEL slot (blocks future jobs / resume compute mode * VM). The caveat is all dma-resv slots must be updated atomically with respect * to execs and compute mode rebind worker. To accomplish this, hold the * vm->lock in write mode from the first operation until the last. * * Deferred binds in fault mode * ---------------------------- * * In a VM is in fault mode (TODO: link to fault mode), new bind operations that * create mappings are by default are deferred to the page fault handler (first * use). This behavior can be overriden by setting the flag * DRM_XE_VM_BIND_FLAG_IMMEDIATE which indicates to creating the mapping * immediately. * * User pointer * ============ * * User pointers are user allocated memory (malloc'd, mmap'd, etc..) for which the * user wants to create a GPU mapping. Typically in other DRM drivers a dummy BO * was created and then a binding was created. We bypass creating a dummy BO in * XE and simply create a binding directly from the userptr. * * Invalidation * ------------ * * Since this a core kernel managed memory the kernel can move this memory * whenever it wants. We register an invalidation MMU notifier to alert XE when * a user poiter is about to move. The invalidation notifier needs to block * until all pending users (jobs or compute mode engines) of the userptr are * idle to ensure no faults. This done by waiting on all of VM's dma-resv slots. * * Rebinds * ------- * * Either the next exec (non-compute) or rebind worker (compute mode) will * rebind the userptr. The invalidation MMU notifier kicks the rebind worker * after the VM dma-resv wait if the VM is in compute mode. * * Compute mode * ============ * * A VM in compute mode enables long running workloads and ultra low latency * submission (ULLS). ULLS is implemented via a continuously running batch + * semaphores. This enables to the user to insert jump to new batch commands * into the continuously running batch. In both cases these batches exceed the * time a dma fence is allowed to exist for before signaling, as such dma fences * are not used when a VM is in compute mode. User fences (TODO: link user fence * doc) are used instead to signal operation's completion. * * Preempt fences * -------------- * * If the kernel decides to move memory around (either userptr invalidate, BO * eviction, or mumap style unbind which results in a rebind) and a batch is * running on an engine, that batch can fault or cause a memory corruption as * page tables for the moved memory are no longer valid. To work around this we * introduce the concept of preempt fences. When sw signaling is enabled on a * preempt fence it tells the submission backend to kick that engine off the * hardware and the preempt fence signals when the engine is off the hardware. * Once all preempt fences are signaled for a VM the kernel can safely move the * memory and kick the rebind worker which resumes all the engines execution. * * A preempt fence, for every engine using the VM, is installed the VM's * dma-resv DMA_RESV_USAGE_PREEMPT_FENCE slot. The same preempt fence, for every * engine using the VM, is also installed into the same dma-resv slot of every * external BO mapped in the VM. * * Rebind worker * ------------- * * The rebind worker is very similar to an exec. It is resposible for rebinding * evicted BOs or userptrs, waiting on those operations, installing new preempt * fences, and finally resuming executing of engines in the VM. * * Flow * ~~~~ * * .. code-block:: * * <----------------------------------------------------------------------| * Check if VM is closed, if so bail out | * Lock VM global lock in read mode | * Pin userptrs (also finds userptr invalidated since last rebind worker) | * Lock VM dma-resv and external BOs dma-resv | * Validate BOs that have been evicted | * Wait on and allocate new preempt fences for every engine using the VM | * Rebind invalidated userptrs + evicted BOs | * Wait on last rebind fence | * Wait VM's DMA_RESV_USAGE_KERNEL dma-resv slot | * Install preeempt fences and issue resume for every engine using the VM | * Check if any userptrs invalidated since pin | * Squash resume for all engines | * Unlock all | * Wait all VM's dma-resv slots | * Retry ---------------------------------------------------------- * Release all engines waiting to resume * Unlock all * * Timeslicing * ----------- * * In order to prevent an engine from continuously being kicked off the hardware * and making no forward progress an engine has a period of time it allowed to * run after resume before it can be kicked off again. This effectively gives * each engine a timeslice. * * Handling multiple GTs * ===================== * * If a GT has slower access to some regions and the page table structure are in * the slow region, the performance on that GT could adversely be affected. To * work around this we allow a VM page tables to be shadowed in multiple GTs. * When VM is created, a default bind engine and PT table structure are created * on each GT. * * Binds can optionally pass in a mask of GTs where a mapping should be created, * if this mask is zero then default to all the GTs where the VM has page * tables. * * The implementation for this breaks down into a bunch for_each_gt loops in * various places plus exporting a composite fence for multi-GT binds to the * user. * * Fault mode (unified shared memory) * ================================== * * A VM in fault mode can be enabled on devices that support page faults. If * page faults are enabled, using dma fences can potentially induce a deadlock: * A pending page fault can hold up the GPU work which holds up the dma fence * signaling, and memory allocation is usually required to resolve a page * fault, but memory allocation is not allowed to gate dma fence signaling. As * such, dma fences are not allowed when VM is in fault mode. Because dma-fences * are not allowed, long running workloads and ULLS are enabled on a faulting * VM. * * Defered VM binds * ---------------- * * By default, on a faulting VM binds just allocate the VMA and the actual * updating of the page tables is defered to the page fault handler. This * behavior can be overridden by setting the flag DRM_XE_VM_BIND_FLAG_IMMEDIATE in * the VM bind which will then do the bind immediately. * * Page fault handler * ------------------ * * Page faults are received in the G2H worker under the CT lock which is in the * path of dma fences (no memory allocations are allowed, faults require memory * allocations) thus we cannot process faults under the CT lock. Another issue * is faults issue TLB invalidations which require G2H credits and we cannot * allocate G2H credits in the G2H handlers without deadlocking. Lastly, we do * not want the CT lock to be an outer lock of the VM global lock (VM global * lock required to fault processing). * * To work around the above issue with processing faults in the G2H worker, we * sink faults to a buffer which is large enough to sink all possible faults on * the GT (1 per hardware engine) and kick a worker to process the faults. Since * the page faults G2H are already received in a worker, kicking another worker * adds more latency to a critical performance path. We add a fast path in the * G2H irq handler which looks at first G2H and if it is a page fault we sink * the fault to the buffer and kick the worker to process the fault. TLB * invalidation responses are also in the critical path so these can also be * processed in this fast path. * * Multiple buffers and workers are used and hashed over based on the ASID so * faults from different VMs can be processed in parallel. * * The page fault handler itself is rather simple, flow is below. * * .. code-block:: * * Lookup VM from ASID in page fault G2H * Lock VM global lock in read mode * Lookup VMA from address in page fault G2H * Check if VMA is valid, if not bail * Check if VMA's BO has backing store, if not allocate * <----------------------------------------------------------------------| * If userptr, pin pages | * Lock VM & BO dma-resv locks | * If atomic fault, migrate to VRAM, else validate BO location | * Issue rebind | * Wait on rebind to complete | * Check if userptr invalidated since pin | * Drop VM & BO dma-resv locks | * Retry ---------------------------------------------------------- * Unlock all * Issue blocking TLB invalidation | * Send page fault response to GuC * * Access counters * --------------- * * Access counters can be configured to trigger a G2H indicating the device is * accessing VMAs in system memory frequently as hint to migrate those VMAs to * VRAM. * * Same as the page fault handler, access counters G2H cannot be processed the * G2H worker under the CT lock. Again we use a buffer to sink access counter * G2H. Unlike page faults there is no upper bound so if the buffer is full we * simply drop the G2H. Access counters are a best case optimization and it is * safe to drop these unlike page faults. * * The access counter handler itself is rather simple flow is below. * * .. code-block:: * * Lookup VM from ASID in access counter G2H * Lock VM global lock in read mode * Lookup VMA from address in access counter G2H * If userptr, bail nothing to do * Lock VM & BO dma-resv locks * Issue migration to VRAM * Unlock all * * Notice no rebind is issued in the access counter handler as the rebind will * be issued on next page fault. * * Cavets with eviction / user pointer invalidation * ------------------------------------------------ * * In the case of eviction and user pointer invalidation on a faulting VM, there * is no need to issue a rebind rather we just need to blow away the page tables * for the VMAs and the page fault handler will rebind the VMAs when they fault. * The cavet is to update / read the page table structure the VM global lock is * neeeed. In both the case of eviction and user pointer invalidation locks are * held which make acquiring the VM global lock impossible. To work around this * every VMA maintains a list of leaf page table entries which should be written * to zero to blow away the VMA's page tables. After writing zero to these * entries a blocking TLB invalidate is issued. At this point it is safe for the * kernel to move the VMA's memory around. This is a necessary lockless * algorithm and is safe as leafs cannot be changed while either an eviction or * userptr invalidation is occurring. * * Locking * ======= * * VM locking protects all of the core data paths (bind operations, execs, * evictions, and compute mode rebind worker) in XE. * * Locks * ----- * * VM global lock (vm->lock) - rw semaphore lock. Outer most lock which protects * the list of userptrs mapped in the VM, the list of engines using this VM, and * the array of external BOs mapped in the VM. When adding or removing any of the * aforemented state from the VM should acquire this lock in write mode. The VM * bind path also acquires this lock in write while the exec / compute mode * rebind worker acquire this lock in read mode. * * VM dma-resv lock (vm->ttm.base.resv->lock) - WW lock. Protects VM dma-resv * slots which is shared with any private BO in the VM. Expected to be acquired * during VM binds, execs, and compute mode rebind worker. This lock is also * held when private BOs are being evicted. * * external BO dma-resv lock (bo->ttm.base.resv->lock) - WW lock. Protects * external BO dma-resv slots. Expected to be acquired during VM binds (in * addition to the VM dma-resv lock). All external BO dma-locks within a VM are * expected to be acquired (in addition to the VM dma-resv lock) during execs * and the compute mode rebind worker. This lock is also held when an external * BO is being evicted. * * Putting it all together * ----------------------- * * 1. An exec and bind operation with the same VM can't be executing at the same * time (vm->lock). * * 2. A compute mode rebind worker and bind operation with the same VM can't be * executing at the same time (vm->lock). * * 3. We can't add / remove userptrs or external BOs to a VM while an exec with * the same VM is executing (vm->lock). * * 4. We can't add / remove userptrs, external BOs, or engines to a VM while a * compute mode rebind worker with the same VM is executing (vm->lock). * * 5. Evictions within a VM can't be happen while an exec with the same VM is * executing (dma-resv locks). * * 6. Evictions within a VM can't be happen while a compute mode rebind worker * with the same VM is executing (dma-resv locks). * * dma-resv usage * ============== * * As previously stated to enforce the ordering of kernel ops (eviction, userptr * invalidation, munmap style unbinds which result in a rebind), rebinds during * execs, execs, and resumes in the rebind worker we use both the VMs and * external BOs dma-resv slots. Let try to make this as clear as possible. * * Slot installation * ----------------- * * 1. Jobs from kernel ops install themselves into the DMA_RESV_USAGE_KERNEL * slot of either an external BO or VM (depends on if kernel op is operating on * an external or private BO) * * 2. In non-compute mode, jobs from execs install themselves into the * DMA_RESV_USAGE_BOOKKEEP slot of the VM * * 3. In non-compute mode, jobs from execs install themselves into the * DMA_RESV_USAGE_WRITE slot of all external BOs in the VM * * 4. Jobs from binds install themselves into the DMA_RESV_USAGE_BOOKKEEP slot * of the VM * * 5. Jobs from binds install themselves into the DMA_RESV_USAGE_BOOKKEEP slot * of the external BO (if the bind is to an external BO, this is addition to #4) * * 6. Every engine using a compute mode VM has a preempt fence in installed into * the DMA_RESV_USAGE_PREEMPT_FENCE slot of the VM * * 7. Every engine using a compute mode VM has a preempt fence in installed into * the DMA_RESV_USAGE_PREEMPT_FENCE slot of all the external BOs in the VM * * Slot waiting * ------------ * * 1. The exection of all jobs from kernel ops shall wait on all slots * (DMA_RESV_USAGE_PREEMPT_FENCE) of either an external BO or VM (depends on if * kernel op is operating on external or private BO) * * 2. In non-compute mode, the exection of all jobs from rebinds in execs shall * wait on the DMA_RESV_USAGE_KERNEL slot of either an external BO or VM * (depends on if the rebind is operatiing on an external or private BO) * * 3. In non-compute mode, the exection of all jobs from execs shall wait on the * last rebind job * * 4. In compute mode, the exection of all jobs from rebinds in the rebind * worker shall wait on the DMA_RESV_USAGE_KERNEL slot of either an external BO * or VM (depends on if rebind is operating on external or private BO) * * 5. In compute mode, resumes in rebind worker shall wait on last rebind fence * * 6. In compute mode, resumes in rebind worker shall wait on the * DMA_RESV_USAGE_KERNEL slot of the VM * * Putting it all together * ----------------------- * * 1. New jobs from kernel ops are blocked behind any existing jobs from * non-compute mode execs * * 2. New jobs from non-compute mode execs are blocked behind any existing jobs * from kernel ops and rebinds * * 3. New jobs from kernel ops are blocked behind all preempt fences signaling in * compute mode * * 4. Compute mode engine resumes are blocked behind any existing jobs from * kernel ops and rebinds * * Future work * =========== * * Support large pages for sysmem and userptr. * * Update page faults to handle BOs are page level grainularity (e.g. part of BO * could be in system memory while another part could be in VRAM). * * Page fault handler likely we be optimized a bit more (e.g. Rebinds always * wait on the dma-resv kernel slots of VM or BO, technically we only have to * wait the BO moving. If using a job to do the rebind, we could not block in * the page fault handler rather attach a callback to fence of the rebind job to * signal page fault complete. Our handling of short circuting for atomic faults * for bound VMAs could be better. etc...). We can tune all of this once we have * benchmarks / performance number from workloads up and running. */ #endif