xe_sched_job.c

  1// SPDX-License-Identifier: MIT
  2/*
  3 * Copyright © 2021 Intel Corporation
  4 */
  5
  6#include "xe_sched_job.h"
  7
  8#include <uapi/drm/xe_drm.h>
  9#include <linux/dma-fence-chain.h>
 10#include <linux/slab.h>
 11
 12#include "xe_device.h"
 13#include "xe_exec_queue.h"
 14#include "xe_gt.h"
 15#include "xe_hw_engine_types.h"
 16#include "xe_hw_fence.h"
 17#include "xe_lrc.h"
 18#include "xe_macros.h"
 19#include "xe_pm.h"
 20#include "xe_sync_types.h"
 21#include "xe_trace.h"
 22#include "xe_vm.h"
 23
 24static struct kmem_cache *xe_sched_job_slab;
 25static struct kmem_cache *xe_sched_job_parallel_slab;
 26
 27int __init xe_sched_job_module_init(void)
 28{
 29	xe_sched_job_slab =
 30		kmem_cache_create("xe_sched_job",
 31				  sizeof(struct xe_sched_job) +
 32				  sizeof(struct xe_job_ptrs), 0,
 33				  SLAB_HWCACHE_ALIGN, NULL);
 34	if (!xe_sched_job_slab)
 35		return -ENOMEM;
 36
 37	xe_sched_job_parallel_slab =
 38		kmem_cache_create("xe_sched_job_parallel",
 39				  sizeof(struct xe_sched_job) +
 40				  sizeof(struct xe_job_ptrs) *
 41				  XE_HW_ENGINE_MAX_INSTANCE, 0,
 42				  SLAB_HWCACHE_ALIGN, NULL);
 43	if (!xe_sched_job_parallel_slab) {
 44		kmem_cache_destroy(xe_sched_job_slab);
 45		return -ENOMEM;
 46	}
 47
 48	return 0;
 49}
 50
 51void xe_sched_job_module_exit(void)
 52{
 53	kmem_cache_destroy(xe_sched_job_slab);
 54	kmem_cache_destroy(xe_sched_job_parallel_slab);
 55}
 56
 57static struct xe_sched_job *job_alloc(bool parallel)
 58{
 59	return kmem_cache_zalloc(parallel ? xe_sched_job_parallel_slab :
 60				 xe_sched_job_slab, GFP_KERNEL);
 61}
 62
 63bool xe_sched_job_is_migration(struct xe_exec_queue *q)
 64{
 65	return q->vm && (q->vm->flags & XE_VM_FLAG_MIGRATION);
 66}
 67
 68static void job_free(struct xe_sched_job *job)
 69{
 70	struct xe_exec_queue *q = job->q;
 71	bool is_migration = xe_sched_job_is_migration(q);
 72
 73	kmem_cache_free(xe_exec_queue_is_parallel(job->q) || is_migration ?
 74			xe_sched_job_parallel_slab : xe_sched_job_slab, job);
 75}
 76
 77static struct xe_device *job_to_xe(struct xe_sched_job *job)
 78{
 79	return gt_to_xe(job->q->gt);
 80}
 81
 82/* Free unused pre-allocated fences */
 83static void xe_sched_job_free_fences(struct xe_sched_job *job)
 84{
 85	int i;
 86
 87	for (i = 0; i < job->q->width; ++i) {
 88		struct xe_job_ptrs *ptrs = &job->ptrs[i];
 89
 90		if (ptrs->lrc_fence)
 91			xe_lrc_free_seqno_fence(ptrs->lrc_fence);
 92		dma_fence_chain_free(ptrs->chain_fence);
 93	}
 94}
 95
 96struct xe_sched_job *xe_sched_job_create(struct xe_exec_queue *q,
 97					 u64 *batch_addr)
 98{
 99	bool is_migration = xe_sched_job_is_migration(q);
100	struct xe_sched_job *job;
101	int err;
102	int i;
103	u32 width;
104
105	/* only a kernel context can submit a vm-less job */
106	XE_WARN_ON(!q->vm && !(q->flags & EXEC_QUEUE_FLAG_KERNEL));
107
108	job = job_alloc(xe_exec_queue_is_parallel(q) || is_migration);
109	if (!job)
110		return ERR_PTR(-ENOMEM);
111
112	job->q = q;
113	kref_init(&job->refcount);
114	xe_exec_queue_get(job->q);
115
116	err = drm_sched_job_init(&job->drm, q->entity, 1, NULL,
117				 q->xef ? q->xef->drm->client_id : 0);
118	if (err)
119		goto err_free;
120
121	for (i = 0; i < q->width; ++i) {
122		struct dma_fence *fence = xe_lrc_alloc_seqno_fence();
123		struct dma_fence_chain *chain;
124
125		if (IS_ERR(fence)) {
126			err = PTR_ERR(fence);
127			goto err_sched_job;
128		}
129		job->ptrs[i].lrc_fence = fence;
130
131		if (i + 1 == q->width)
132			continue;
133
134		chain = dma_fence_chain_alloc();
135		if (!chain) {
136			err = -ENOMEM;
137			goto err_sched_job;
138		}
139		job->ptrs[i].chain_fence = chain;
140	}
141
142	width = q->width;
143	if (is_migration)
144		width = 2;
145
146	for (i = 0; i < width; ++i)
147		job->ptrs[i].batch_addr = batch_addr[i];
148
149	atomic_inc(&q->job_cnt);
150	xe_pm_runtime_get_noresume(job_to_xe(job));
151	trace_xe_sched_job_create(job);
152	return job;
153
154err_sched_job:
155	xe_sched_job_free_fences(job);
156	drm_sched_job_cleanup(&job->drm);
157err_free:
158	xe_exec_queue_put(q);
159	job_free(job);
160	return ERR_PTR(err);
161}
162
163/**
164 * xe_sched_job_destroy - Destroy Xe schedule job
165 * @ref: reference to Xe schedule job
166 *
167 * Called when ref == 0, drop a reference to job's xe_engine + fence, cleanup
168 * base DRM schedule job, and free memory for Xe schedule job.
169 */
170void xe_sched_job_destroy(struct kref *ref)
171{
172	struct xe_sched_job *job =
173		container_of(ref, struct xe_sched_job, refcount);
174	struct xe_device *xe = job_to_xe(job);
175	struct xe_exec_queue *q = job->q;
176
177	xe_sched_job_free_fences(job);
178	dma_fence_put(job->fence);
179	drm_sched_job_cleanup(&job->drm);
180	job_free(job);
181	atomic_dec(&q->job_cnt);
182	xe_exec_queue_put(q);
183	xe_pm_runtime_put(xe);
184}
185
186/* Set the error status under the fence to avoid racing with signaling */
187static bool xe_fence_set_error(struct dma_fence *fence, int error)
188{
189	unsigned long irq_flags;
190	bool signaled;
191
192	spin_lock_irqsave(fence->lock, irq_flags);
193	signaled = test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags);
194	if (!signaled)
195		dma_fence_set_error(fence, error);
196	spin_unlock_irqrestore(fence->lock, irq_flags);
197
198	return signaled;
199}
200
201void xe_sched_job_set_error(struct xe_sched_job *job, int error)
202{
203	if (xe_fence_set_error(job->fence, error))
204		return;
205
206	if (dma_fence_is_chain(job->fence)) {
207		struct dma_fence *iter;
208
209		dma_fence_chain_for_each(iter, job->fence)
210			xe_fence_set_error(dma_fence_chain_contained(iter),
211					   error);
212	}
213
214	trace_xe_sched_job_set_error(job);
215
216	dma_fence_enable_sw_signaling(job->fence);
217	xe_hw_fence_irq_run(job->q->fence_irq);
218}
219
220bool xe_sched_job_started(struct xe_sched_job *job)
221{
222	struct dma_fence *fence = dma_fence_chain_contained(job->fence);
223	struct xe_lrc *lrc = job->q->lrc[0];
224
225	return !__dma_fence_is_later(fence,
226				     xe_sched_job_lrc_seqno(job),
227				     xe_lrc_start_seqno(lrc));
228}
229
230bool xe_sched_job_completed(struct xe_sched_job *job)
231{
232	struct dma_fence *fence = dma_fence_chain_contained(job->fence);
233	struct xe_lrc *lrc = job->q->lrc[0];
234
235	/*
236	 * Can safely check just LRC[0] seqno as that is last seqno written when
237	 * parallel handshake is done.
238	 */
239
240	return !__dma_fence_is_later(fence,
241				     xe_sched_job_lrc_seqno(job),
242				     xe_lrc_seqno(lrc));
243}
244
245void xe_sched_job_arm(struct xe_sched_job *job)
246{
247	struct xe_exec_queue *q = job->q;
248	struct dma_fence *fence, *prev;
249	struct xe_vm *vm = q->vm;
250	u64 seqno = 0;
251	int i;
252
253	/* Migration and kernel engines have their own locking */
254	if (IS_ENABLED(CONFIG_LOCKDEP) &&
255	    !(q->flags & (EXEC_QUEUE_FLAG_KERNEL | EXEC_QUEUE_FLAG_VM))) {
256		lockdep_assert_held(&q->vm->lock);
257		if (!xe_vm_in_lr_mode(q->vm))
258			xe_vm_assert_held(q->vm);
259	}
260
261	if (vm && !xe_sched_job_is_migration(q) && !xe_vm_in_lr_mode(vm) &&
262	    (vm->batch_invalidate_tlb || vm->tlb_flush_seqno != q->tlb_flush_seqno)) {
263		xe_vm_assert_held(vm);
264		q->tlb_flush_seqno = vm->tlb_flush_seqno;
265		job->ring_ops_flush_tlb = true;
266	}
267
268	/* Arm the pre-allocated fences */
269	for (i = 0; i < q->width; prev = fence, ++i) {
270		struct dma_fence_chain *chain;
271
272		fence = job->ptrs[i].lrc_fence;
273		xe_lrc_init_seqno_fence(q->lrc[i], fence);
274		job->ptrs[i].lrc_fence = NULL;
275		if (!i) {
276			job->lrc_seqno = fence->seqno;
277			continue;
278		} else {
279			xe_assert(gt_to_xe(q->gt), job->lrc_seqno == fence->seqno);
280		}
281
282		chain = job->ptrs[i - 1].chain_fence;
283		dma_fence_chain_init(chain, prev, fence, seqno++);
284		job->ptrs[i - 1].chain_fence = NULL;
285		fence = &chain->base;
286	}
287
288	job->fence = dma_fence_get(fence);	/* Pairs with put in scheduler */
289	drm_sched_job_arm(&job->drm);
290}
291
292void xe_sched_job_push(struct xe_sched_job *job)
293{
294	xe_sched_job_get(job);
295	trace_xe_sched_job_exec(job);
296	drm_sched_entity_push_job(&job->drm);
297	xe_sched_job_put(job);
298}
299
300/**
301 * xe_sched_job_init_user_fence - Initialize user_fence for the job
302 * @job: job whose user_fence needs an init
303 * @sync: sync to be use to init user_fence
304 */
305void xe_sched_job_init_user_fence(struct xe_sched_job *job,
306				  struct xe_sync_entry *sync)
307{
308	if (sync->type != DRM_XE_SYNC_TYPE_USER_FENCE)
309		return;
310
311	job->user_fence.used = true;
312	job->user_fence.addr = sync->addr;
313	job->user_fence.value = sync->timeline_value;
314}
315
316struct xe_sched_job_snapshot *
317xe_sched_job_snapshot_capture(struct xe_sched_job *job)
318{
319	struct xe_exec_queue *q = job->q;
320	struct xe_device *xe = q->gt->tile->xe;
321	struct xe_sched_job_snapshot *snapshot;
322	size_t len = sizeof(*snapshot) + (sizeof(u64) * q->width);
323	u16 i;
324
325	snapshot = kzalloc(len, GFP_ATOMIC);
326	if (!snapshot)
327		return NULL;
328
329	snapshot->batch_addr_len = q->width;
330	for (i = 0; i < q->width; i++)
331		snapshot->batch_addr[i] =
332			xe_device_uncanonicalize_addr(xe, job->ptrs[i].batch_addr);
333
334	return snapshot;
335}
336
337void xe_sched_job_snapshot_free(struct xe_sched_job_snapshot *snapshot)
338{
339	kfree(snapshot);
340}
341
342void
343xe_sched_job_snapshot_print(struct xe_sched_job_snapshot *snapshot,
344			    struct drm_printer *p)
345{
346	u16 i;
347
348	if (!snapshot)
349		return;
350
351	for (i = 0; i < snapshot->batch_addr_len; i++)
352		drm_printf(p, "batch_addr[%u]: 0x%016llx\n", i, snapshot->batch_addr[i]);
353}
354
355int xe_sched_job_add_deps(struct xe_sched_job *job, struct dma_resv *resv,
356			  enum dma_resv_usage usage)
357{
358	return drm_sched_job_add_resv_dependencies(&job->drm, resv, usage);
359}