brintos

brintos / linux-shallow public Read only

0
0
Text · 16.3 KiB · 92d27d3 Raw
638 lines · c
1/*2 * Copyright 2017 Advanced Micro Devices, Inc.3 *4 * Permission is hereby granted, free of charge, to any person obtaining a5 * copy of this software and associated documentation files (the "Software"),6 * to deal in the Software without restriction, including without limitation7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,8 * and/or sell copies of the Software, and to permit persons to whom the9 * Software is furnished to do so, subject to the following conditions:10 *11 * The above copyright notice and this permission notice shall be included in12 * all copies or substantial portions of the Software.13 *14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR20 * OTHER DEALINGS IN THE SOFTWARE.21 *22 */23#include "amdgpu_ids.h"24 25#include <linux/idr.h>26#include <linux/dma-fence-array.h>27 28 29#include "amdgpu.h"30#include "amdgpu_trace.h"31 32/*33 * PASID manager34 *35 * PASIDs are global address space identifiers that can be shared36 * between the GPU, an IOMMU and the driver. VMs on different devices37 * may use the same PASID if they share the same address38 * space. Therefore PASIDs are allocated using a global IDA. VMs are39 * looked up from the PASID per amdgpu_device.40 */41static DEFINE_IDA(amdgpu_pasid_ida);42 43/* Helper to free pasid from a fence callback */44struct amdgpu_pasid_cb {45	struct dma_fence_cb cb;46	u32 pasid;47};48 49/**50 * amdgpu_pasid_alloc - Allocate a PASID51 * @bits: Maximum width of the PASID in bits, must be at least 152 *53 * Allocates a PASID of the given width while keeping smaller PASIDs54 * available if possible.55 *56 * Returns a positive integer on success. Returns %-EINVAL if bits==0.57 * Returns %-ENOSPC if no PASID was available. Returns %-ENOMEM on58 * memory allocation failure.59 */60int amdgpu_pasid_alloc(unsigned int bits)61{62	int pasid = -EINVAL;63 64	for (bits = min(bits, 31U); bits > 0; bits--) {65		pasid = ida_alloc_range(&amdgpu_pasid_ida, 1U << (bits - 1),66					(1U << bits) - 1, GFP_KERNEL);67		if (pasid != -ENOSPC)68			break;69	}70 71	if (pasid >= 0)72		trace_amdgpu_pasid_allocated(pasid);73 74	return pasid;75}76 77/**78 * amdgpu_pasid_free - Free a PASID79 * @pasid: PASID to free80 */81void amdgpu_pasid_free(u32 pasid)82{83	trace_amdgpu_pasid_freed(pasid);84	ida_free(&amdgpu_pasid_ida, pasid);85}86 87static void amdgpu_pasid_free_cb(struct dma_fence *fence,88				 struct dma_fence_cb *_cb)89{90	struct amdgpu_pasid_cb *cb =91		container_of(_cb, struct amdgpu_pasid_cb, cb);92 93	amdgpu_pasid_free(cb->pasid);94	dma_fence_put(fence);95	kfree(cb);96}97 98/**99 * amdgpu_pasid_free_delayed - free pasid when fences signal100 *101 * @resv: reservation object with the fences to wait for102 * @pasid: pasid to free103 *104 * Free the pasid only after all the fences in resv are signaled.105 */106void amdgpu_pasid_free_delayed(struct dma_resv *resv,107			       u32 pasid)108{109	struct amdgpu_pasid_cb *cb;110	struct dma_fence *fence;111	int r;112 113	r = dma_resv_get_singleton(resv, DMA_RESV_USAGE_BOOKKEEP, &fence);114	if (r)115		goto fallback;116 117	if (!fence) {118		amdgpu_pasid_free(pasid);119		return;120	}121 122	cb = kmalloc(sizeof(*cb), GFP_KERNEL);123	if (!cb) {124		/* Last resort when we are OOM */125		dma_fence_wait(fence, false);126		dma_fence_put(fence);127		amdgpu_pasid_free(pasid);128	} else {129		cb->pasid = pasid;130		if (dma_fence_add_callback(fence, &cb->cb,131					   amdgpu_pasid_free_cb))132			amdgpu_pasid_free_cb(fence, &cb->cb);133	}134 135	return;136 137fallback:138	/* Not enough memory for the delayed delete, as last resort139	 * block for all the fences to complete.140	 */141	dma_resv_wait_timeout(resv, DMA_RESV_USAGE_BOOKKEEP,142			      false, MAX_SCHEDULE_TIMEOUT);143	amdgpu_pasid_free(pasid);144}145 146/*147 * VMID manager148 *149 * VMIDs are a per VMHUB identifier for page tables handling.150 */151 152/**153 * amdgpu_vmid_had_gpu_reset - check if reset occured since last use154 *155 * @adev: amdgpu_device pointer156 * @id: VMID structure157 *158 * Check if GPU reset occured since last use of the VMID.159 */160bool amdgpu_vmid_had_gpu_reset(struct amdgpu_device *adev,161			       struct amdgpu_vmid *id)162{163	return id->current_gpu_reset_count !=164		atomic_read(&adev->gpu_reset_counter);165}166 167/* Check if we need to switch to another set of resources */168static bool amdgpu_vmid_gds_switch_needed(struct amdgpu_vmid *id,169					  struct amdgpu_job *job)170{171	return id->gds_base != job->gds_base ||172		id->gds_size != job->gds_size ||173		id->gws_base != job->gws_base ||174		id->gws_size != job->gws_size ||175		id->oa_base != job->oa_base ||176		id->oa_size != job->oa_size;177}178 179/* Check if the id is compatible with the job */180static bool amdgpu_vmid_compatible(struct amdgpu_vmid *id,181				   struct amdgpu_job *job)182{183	return  id->pd_gpu_addr == job->vm_pd_addr &&184		!amdgpu_vmid_gds_switch_needed(id, job);185}186 187/**188 * amdgpu_vmid_grab_idle - grab idle VMID189 *190 * @ring: ring we want to submit job to191 * @idle: resulting idle VMID192 * @fence: fence to wait for if no id could be grabbed193 *194 * Try to find an idle VMID, if none is idle add a fence to wait to the sync195 * object. Returns -ENOMEM when we are out of memory.196 */197static int amdgpu_vmid_grab_idle(struct amdgpu_ring *ring,198				 struct amdgpu_vmid **idle,199				 struct dma_fence **fence)200{201	struct amdgpu_device *adev = ring->adev;202	unsigned vmhub = ring->vm_hub;203	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];204	struct dma_fence **fences;205	unsigned i;206 207	if (!dma_fence_is_signaled(ring->vmid_wait)) {208		*fence = dma_fence_get(ring->vmid_wait);209		return 0;210	}211 212	fences = kmalloc_array(id_mgr->num_ids, sizeof(void *), GFP_KERNEL);213	if (!fences)214		return -ENOMEM;215 216	/* Check if we have an idle VMID */217	i = 0;218	list_for_each_entry((*idle), &id_mgr->ids_lru, list) {219		/* Don't use per engine and per process VMID at the same time */220		struct amdgpu_ring *r = adev->vm_manager.concurrent_flush ?221			NULL : ring;222 223		fences[i] = amdgpu_sync_peek_fence(&(*idle)->active, r);224		if (!fences[i])225			break;226		++i;227	}228 229	/* If we can't find a idle VMID to use, wait till one becomes available */230	if (&(*idle)->list == &id_mgr->ids_lru) {231		u64 fence_context = adev->vm_manager.fence_context + ring->idx;232		unsigned seqno = ++adev->vm_manager.seqno[ring->idx];233		struct dma_fence_array *array;234		unsigned j;235 236		*idle = NULL;237		for (j = 0; j < i; ++j)238			dma_fence_get(fences[j]);239 240		array = dma_fence_array_create(i, fences, fence_context,241					       seqno, true);242		if (!array) {243			for (j = 0; j < i; ++j)244				dma_fence_put(fences[j]);245			kfree(fences);246			return -ENOMEM;247		}248 249		*fence = dma_fence_get(&array->base);250		dma_fence_put(ring->vmid_wait);251		ring->vmid_wait = &array->base;252		return 0;253	}254	kfree(fences);255 256	return 0;257}258 259/**260 * amdgpu_vmid_grab_reserved - try to assign reserved VMID261 *262 * @vm: vm to allocate id for263 * @ring: ring we want to submit job to264 * @job: job who wants to use the VMID265 * @id: resulting VMID266 * @fence: fence to wait for if no id could be grabbed267 *268 * Try to assign a reserved VMID.269 */270static int amdgpu_vmid_grab_reserved(struct amdgpu_vm *vm,271				     struct amdgpu_ring *ring,272				     struct amdgpu_job *job,273				     struct amdgpu_vmid **id,274				     struct dma_fence **fence)275{276	struct amdgpu_device *adev = ring->adev;277	unsigned vmhub = ring->vm_hub;278	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];279	uint64_t fence_context = adev->fence_context + ring->idx;280	bool needs_flush = vm->use_cpu_for_update;281	uint64_t updates = amdgpu_vm_tlb_seq(vm);282	int r;283 284	*id = id_mgr->reserved;285	if ((*id)->owner != vm->immediate.fence_context ||286	    !amdgpu_vmid_compatible(*id, job) ||287	    (*id)->flushed_updates < updates ||288	    !(*id)->last_flush ||289	    ((*id)->last_flush->context != fence_context &&290	     !dma_fence_is_signaled((*id)->last_flush))) {291		struct dma_fence *tmp;292 293		/* Wait for the gang to be assembled before using a294		 * reserved VMID or otherwise the gang could deadlock.295		 */296		tmp = amdgpu_device_get_gang(adev);297		if (!dma_fence_is_signaled(tmp) && tmp != job->gang_submit) {298			*id = NULL;299			*fence = tmp;300			return 0;301		}302		dma_fence_put(tmp);303 304		/* Make sure the id is owned by the gang before proceeding */305		if (!job->gang_submit ||306		    (*id)->owner != vm->immediate.fence_context) {307 308			/* Don't use per engine and per process VMID at the309			 * same time310			 */311			if (adev->vm_manager.concurrent_flush)312				ring = NULL;313 314			/* to prevent one context starved by another context */315			(*id)->pd_gpu_addr = 0;316			tmp = amdgpu_sync_peek_fence(&(*id)->active, ring);317			if (tmp) {318				*id = NULL;319				*fence = dma_fence_get(tmp);320				return 0;321			}322		}323		needs_flush = true;324	}325 326	/* Good we can use this VMID. Remember this submission as327	* user of the VMID.328	*/329	r = amdgpu_sync_fence(&(*id)->active, &job->base.s_fence->finished);330	if (r)331		return r;332 333	job->vm_needs_flush = needs_flush;334	job->spm_update_needed = true;335	return 0;336}337 338/**339 * amdgpu_vmid_grab_used - try to reuse a VMID340 *341 * @vm: vm to allocate id for342 * @ring: ring we want to submit job to343 * @job: job who wants to use the VMID344 * @id: resulting VMID345 * @fence: fence to wait for if no id could be grabbed346 *347 * Try to reuse a VMID for this submission.348 */349static int amdgpu_vmid_grab_used(struct amdgpu_vm *vm,350				 struct amdgpu_ring *ring,351				 struct amdgpu_job *job,352				 struct amdgpu_vmid **id,353				 struct dma_fence **fence)354{355	struct amdgpu_device *adev = ring->adev;356	unsigned vmhub = ring->vm_hub;357	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];358	uint64_t fence_context = adev->fence_context + ring->idx;359	uint64_t updates = amdgpu_vm_tlb_seq(vm);360	int r;361 362	job->vm_needs_flush = vm->use_cpu_for_update;363 364	/* Check if we can use a VMID already assigned to this VM */365	list_for_each_entry_reverse((*id), &id_mgr->ids_lru, list) {366		bool needs_flush = vm->use_cpu_for_update;367 368		/* Check all the prerequisites to using this VMID */369		if ((*id)->owner != vm->immediate.fence_context)370			continue;371 372		if (!amdgpu_vmid_compatible(*id, job))373			continue;374 375		if (!(*id)->last_flush ||376		    ((*id)->last_flush->context != fence_context &&377		     !dma_fence_is_signaled((*id)->last_flush)))378			needs_flush = true;379 380		if ((*id)->flushed_updates < updates)381			needs_flush = true;382 383		if (needs_flush && !adev->vm_manager.concurrent_flush)384			continue;385 386		/* Good, we can use this VMID. Remember this submission as387		 * user of the VMID.388		 */389		r = amdgpu_sync_fence(&(*id)->active,390				      &job->base.s_fence->finished);391		if (r)392			return r;393 394		job->vm_needs_flush |= needs_flush;395		return 0;396	}397 398	*id = NULL;399	return 0;400}401 402/**403 * amdgpu_vmid_grab - allocate the next free VMID404 *405 * @vm: vm to allocate id for406 * @ring: ring we want to submit job to407 * @job: job who wants to use the VMID408 * @fence: fence to wait for if no id could be grabbed409 *410 * Allocate an id for the vm, adding fences to the sync obj as necessary.411 */412int amdgpu_vmid_grab(struct amdgpu_vm *vm, struct amdgpu_ring *ring,413		     struct amdgpu_job *job, struct dma_fence **fence)414{415	struct amdgpu_device *adev = ring->adev;416	unsigned vmhub = ring->vm_hub;417	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];418	struct amdgpu_vmid *idle = NULL;419	struct amdgpu_vmid *id = NULL;420	int r = 0;421 422	mutex_lock(&id_mgr->lock);423	r = amdgpu_vmid_grab_idle(ring, &idle, fence);424	if (r || !idle)425		goto error;426 427	if (amdgpu_vmid_uses_reserved(adev, vm, vmhub)) {428		r = amdgpu_vmid_grab_reserved(vm, ring, job, &id, fence);429		if (r || !id)430			goto error;431	} else {432		r = amdgpu_vmid_grab_used(vm, ring, job, &id, fence);433		if (r)434			goto error;435 436		if (!id) {437			/* Still no ID to use? Then use the idle one found earlier */438			id = idle;439 440			/* Remember this submission as user of the VMID */441			r = amdgpu_sync_fence(&id->active,442					      &job->base.s_fence->finished);443			if (r)444				goto error;445 446			job->vm_needs_flush = true;447		}448 449		list_move_tail(&id->list, &id_mgr->ids_lru);450	}451 452	job->gds_switch_needed = amdgpu_vmid_gds_switch_needed(id, job);453	if (job->vm_needs_flush) {454		id->flushed_updates = amdgpu_vm_tlb_seq(vm);455		dma_fence_put(id->last_flush);456		id->last_flush = NULL;457	}458	job->vmid = id - id_mgr->ids;459	job->pasid = vm->pasid;460 461	id->gds_base = job->gds_base;462	id->gds_size = job->gds_size;463	id->gws_base = job->gws_base;464	id->gws_size = job->gws_size;465	id->oa_base = job->oa_base;466	id->oa_size = job->oa_size;467	id->pd_gpu_addr = job->vm_pd_addr;468	id->owner = vm->immediate.fence_context;469 470	trace_amdgpu_vm_grab_id(vm, ring, job);471 472error:473	mutex_unlock(&id_mgr->lock);474	return r;475}476 477/*478 * amdgpu_vmid_uses_reserved - check if a VM will use a reserved VMID479 * @adev: amdgpu_device pointer480 * @vm: the VM to check481 * @vmhub: the VMHUB which will be used482 *483 * Returns: True if the VM will use a reserved VMID.484 */485bool amdgpu_vmid_uses_reserved(struct amdgpu_device *adev,486			       struct amdgpu_vm *vm, unsigned int vmhub)487{488	return vm->reserved_vmid[vmhub] ||489		(adev->enforce_isolation[(vm->root.bo->xcp_id != AMDGPU_XCP_NO_PARTITION) ?490					 vm->root.bo->xcp_id : 0] &&491		 AMDGPU_IS_GFXHUB(vmhub));492}493 494int amdgpu_vmid_alloc_reserved(struct amdgpu_device *adev,495			       unsigned vmhub)496{497	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];498 499	mutex_lock(&id_mgr->lock);500 501	++id_mgr->reserved_use_count;502	if (!id_mgr->reserved) {503		struct amdgpu_vmid *id;504 505		id = list_first_entry(&id_mgr->ids_lru, struct amdgpu_vmid,506				      list);507		/* Remove from normal round robin handling */508		list_del_init(&id->list);509		id_mgr->reserved = id;510	}511 512	mutex_unlock(&id_mgr->lock);513	return 0;514}515 516void amdgpu_vmid_free_reserved(struct amdgpu_device *adev,517			       unsigned vmhub)518{519	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];520 521	mutex_lock(&id_mgr->lock);522	if (!--id_mgr->reserved_use_count) {523		/* give the reserved ID back to normal round robin */524		list_add(&id_mgr->reserved->list, &id_mgr->ids_lru);525		id_mgr->reserved = NULL;526	}527 528	mutex_unlock(&id_mgr->lock);529}530 531/**532 * amdgpu_vmid_reset - reset VMID to zero533 *534 * @adev: amdgpu device structure535 * @vmhub: vmhub type536 * @vmid: vmid number to use537 *538 * Reset saved GDW, GWS and OA to force switch on next flush.539 */540void amdgpu_vmid_reset(struct amdgpu_device *adev, unsigned vmhub,541		       unsigned vmid)542{543	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];544	struct amdgpu_vmid *id = &id_mgr->ids[vmid];545 546	mutex_lock(&id_mgr->lock);547	id->owner = 0;548	id->gds_base = 0;549	id->gds_size = 0;550	id->gws_base = 0;551	id->gws_size = 0;552	id->oa_base = 0;553	id->oa_size = 0;554	mutex_unlock(&id_mgr->lock);555}556 557/**558 * amdgpu_vmid_reset_all - reset VMID to zero559 *560 * @adev: amdgpu device structure561 *562 * Reset VMID to force flush on next use563 */564void amdgpu_vmid_reset_all(struct amdgpu_device *adev)565{566	unsigned i, j;567 568	for (i = 0; i < AMDGPU_MAX_VMHUBS; ++i) {569		struct amdgpu_vmid_mgr *id_mgr =570			&adev->vm_manager.id_mgr[i];571 572		for (j = 1; j < id_mgr->num_ids; ++j)573			amdgpu_vmid_reset(adev, i, j);574	}575}576 577/**578 * amdgpu_vmid_mgr_init - init the VMID manager579 *580 * @adev: amdgpu_device pointer581 *582 * Initialize the VM manager structures583 */584void amdgpu_vmid_mgr_init(struct amdgpu_device *adev)585{586	unsigned i, j;587 588	for (i = 0; i < AMDGPU_MAX_VMHUBS; ++i) {589		struct amdgpu_vmid_mgr *id_mgr =590			&adev->vm_manager.id_mgr[i];591 592		mutex_init(&id_mgr->lock);593		INIT_LIST_HEAD(&id_mgr->ids_lru);594		id_mgr->reserved_use_count = 0;595 596		/* manage only VMIDs not used by KFD */597		id_mgr->num_ids = adev->vm_manager.first_kfd_vmid;598 599		/* skip over VMID 0, since it is the system VM */600		for (j = 1; j < id_mgr->num_ids; ++j) {601			amdgpu_vmid_reset(adev, i, j);602			amdgpu_sync_create(&id_mgr->ids[j].active);603			list_add_tail(&id_mgr->ids[j].list, &id_mgr->ids_lru);604		}605	}606	/* alloc a default reserved vmid to enforce isolation */607	for (i = 0; i < (adev->xcp_mgr ? adev->xcp_mgr->num_xcps : 1); i++) {608		if (adev->enforce_isolation[i])609			amdgpu_vmid_alloc_reserved(adev, AMDGPU_GFXHUB(i));610	}611}612 613/**614 * amdgpu_vmid_mgr_fini - cleanup VM manager615 *616 * @adev: amdgpu_device pointer617 *618 * Cleanup the VM manager and free resources.619 */620void amdgpu_vmid_mgr_fini(struct amdgpu_device *adev)621{622	unsigned i, j;623 624	for (i = 0; i < AMDGPU_MAX_VMHUBS; ++i) {625		struct amdgpu_vmid_mgr *id_mgr =626			&adev->vm_manager.id_mgr[i];627 628		mutex_destroy(&id_mgr->lock);629		for (j = 0; j < AMDGPU_NUM_VMID; ++j) {630			struct amdgpu_vmid *id = &id_mgr->ids[j];631 632			amdgpu_sync_free(&id->active);633			dma_fence_put(id->last_flush);634			dma_fence_put(id->pasid_mapping);635		}636	}637}638