1 /*
2 * Copyright(c) 2011-2016 Intel Corporation. All rights reserved.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 *
23 * Authors:
24 * Zhi Wang <zhi.a.wang@intel.com>
25 *
26 * Contributors:
27 * Ping Gao <ping.a.gao@intel.com>
28 * Tina Zhang <tina.zhang@intel.com>
29 * Chanbin Du <changbin.du@intel.com>
30 * Min He <min.he@intel.com>
31 * Bing Niu <bing.niu@intel.com>
32 * Zhenyu Wang <zhenyuw@linux.intel.com>
33 *
34 */
35
36 #include <linux/kthread.h>
37
38 #include "gem/i915_gem_pm.h"
39 #include "gt/intel_context.h"
40 #include "gt/intel_execlists_submission.h"
41 #include "gt/intel_gt_regs.h"
42 #include "gt/intel_lrc.h"
43 #include "gt/intel_ring.h"
44
45 #include "i915_drv.h"
46 #include "i915_gem_gtt.h"
47 #include "i915_perf_oa_regs.h"
48 #include "gvt.h"
49
50 #define RING_CTX_OFF(x) \
51 offsetof(struct execlist_ring_context, x)
52
set_context_pdp_root_pointer(struct execlist_ring_context * ring_context,u32 pdp[8])53 static void set_context_pdp_root_pointer(
54 struct execlist_ring_context *ring_context,
55 u32 pdp[8])
56 {
57 int i;
58
59 for (i = 0; i < 8; i++)
60 ring_context->pdps[i].val = pdp[7 - i];
61 }
62
update_shadow_pdps(struct intel_vgpu_workload * workload)63 static void update_shadow_pdps(struct intel_vgpu_workload *workload)
64 {
65 struct execlist_ring_context *shadow_ring_context;
66 struct intel_context *ctx = workload->req->context;
67
68 if (WARN_ON(!workload->shadow_mm))
69 return;
70
71 if (WARN_ON(!atomic_read(&workload->shadow_mm->pincount)))
72 return;
73
74 shadow_ring_context = (struct execlist_ring_context *)ctx->lrc_reg_state;
75 set_context_pdp_root_pointer(shadow_ring_context,
76 (void *)workload->shadow_mm->ppgtt_mm.shadow_pdps);
77 }
78
79 /*
80 * When populating shadow ctx from guest, we should not override oa related
81 * registers, so that they will not be overlapped by guest oa configs. Thus
82 * made it possible to capture oa data from host for both host and guests.
83 */
sr_oa_regs(struct intel_vgpu_workload * workload,u32 * reg_state,bool save)84 static void sr_oa_regs(struct intel_vgpu_workload *workload,
85 u32 *reg_state, bool save)
86 {
87 struct drm_i915_private *dev_priv = workload->vgpu->gvt->gt->i915;
88 u32 ctx_oactxctrl = dev_priv->perf.ctx_oactxctrl_offset;
89 u32 ctx_flexeu0 = dev_priv->perf.ctx_flexeu0_offset;
90 int i = 0;
91 u32 flex_mmio[] = {
92 i915_mmio_reg_offset(EU_PERF_CNTL0),
93 i915_mmio_reg_offset(EU_PERF_CNTL1),
94 i915_mmio_reg_offset(EU_PERF_CNTL2),
95 i915_mmio_reg_offset(EU_PERF_CNTL3),
96 i915_mmio_reg_offset(EU_PERF_CNTL4),
97 i915_mmio_reg_offset(EU_PERF_CNTL5),
98 i915_mmio_reg_offset(EU_PERF_CNTL6),
99 };
100
101 if (workload->engine->id != RCS0)
102 return;
103
104 if (save) {
105 workload->oactxctrl = reg_state[ctx_oactxctrl + 1];
106
107 for (i = 0; i < ARRAY_SIZE(workload->flex_mmio); i++) {
108 u32 state_offset = ctx_flexeu0 + i * 2;
109
110 workload->flex_mmio[i] = reg_state[state_offset + 1];
111 }
112 } else {
113 reg_state[ctx_oactxctrl] =
114 i915_mmio_reg_offset(GEN8_OACTXCONTROL);
115 reg_state[ctx_oactxctrl + 1] = workload->oactxctrl;
116
117 for (i = 0; i < ARRAY_SIZE(workload->flex_mmio); i++) {
118 u32 state_offset = ctx_flexeu0 + i * 2;
119 u32 mmio = flex_mmio[i];
120
121 reg_state[state_offset] = mmio;
122 reg_state[state_offset + 1] = workload->flex_mmio[i];
123 }
124 }
125 }
126
populate_shadow_context(struct intel_vgpu_workload * workload)127 static int populate_shadow_context(struct intel_vgpu_workload *workload)
128 {
129 struct intel_vgpu *vgpu = workload->vgpu;
130 struct intel_gvt *gvt = vgpu->gvt;
131 struct intel_context *ctx = workload->req->context;
132 struct execlist_ring_context *shadow_ring_context;
133 void *dst;
134 void *context_base;
135 unsigned long context_gpa, context_page_num;
136 unsigned long gpa_base; /* first gpa of consecutive GPAs */
137 unsigned long gpa_size; /* size of consecutive GPAs */
138 struct intel_vgpu_submission *s = &vgpu->submission;
139 int i;
140 bool skip = false;
141 int ring_id = workload->engine->id;
142 int ret;
143
144 GEM_BUG_ON(!intel_context_is_pinned(ctx));
145
146 context_base = (void *) ctx->lrc_reg_state -
147 (LRC_STATE_PN << I915_GTT_PAGE_SHIFT);
148
149 shadow_ring_context = (void *) ctx->lrc_reg_state;
150
151 sr_oa_regs(workload, (u32 *)shadow_ring_context, true);
152 #define COPY_REG(name) \
153 intel_gvt_read_gpa(vgpu, workload->ring_context_gpa \
154 + RING_CTX_OFF(name.val), &shadow_ring_context->name.val, 4)
155 #define COPY_REG_MASKED(name) {\
156 intel_gvt_read_gpa(vgpu, workload->ring_context_gpa \
157 + RING_CTX_OFF(name.val),\
158 &shadow_ring_context->name.val, 4);\
159 shadow_ring_context->name.val |= 0xffff << 16;\
160 }
161
162 COPY_REG_MASKED(ctx_ctrl);
163 COPY_REG(ctx_timestamp);
164
165 if (workload->engine->id == RCS0) {
166 COPY_REG(bb_per_ctx_ptr);
167 COPY_REG(rcs_indirect_ctx);
168 COPY_REG(rcs_indirect_ctx_offset);
169 } else if (workload->engine->id == BCS0)
170 intel_gvt_read_gpa(vgpu,
171 workload->ring_context_gpa +
172 BCS_TILE_REGISTER_VAL_OFFSET,
173 (void *)shadow_ring_context +
174 BCS_TILE_REGISTER_VAL_OFFSET, 4);
175 #undef COPY_REG
176 #undef COPY_REG_MASKED
177
178 /* don't copy Ring Context (the first 0x50 dwords),
179 * only copy the Engine Context part from guest
180 */
181 intel_gvt_read_gpa(vgpu,
182 workload->ring_context_gpa +
183 RING_CTX_SIZE,
184 (void *)shadow_ring_context +
185 RING_CTX_SIZE,
186 I915_GTT_PAGE_SIZE - RING_CTX_SIZE);
187
188 sr_oa_regs(workload, (u32 *)shadow_ring_context, false);
189
190 gvt_dbg_sched("ring %s workload lrca %x, ctx_id %x, ctx gpa %llx",
191 workload->engine->name, workload->ctx_desc.lrca,
192 workload->ctx_desc.context_id,
193 workload->ring_context_gpa);
194
195 /* only need to ensure this context is not pinned/unpinned during the
196 * period from last submission to this this submission.
197 * Upon reaching this function, the currently submitted context is not
198 * supposed to get unpinned. If a misbehaving guest driver ever does
199 * this, it would corrupt itself.
200 */
201 if (s->last_ctx[ring_id].valid &&
202 (s->last_ctx[ring_id].lrca ==
203 workload->ctx_desc.lrca) &&
204 (s->last_ctx[ring_id].ring_context_gpa ==
205 workload->ring_context_gpa))
206 skip = true;
207
208 s->last_ctx[ring_id].lrca = workload->ctx_desc.lrca;
209 s->last_ctx[ring_id].ring_context_gpa = workload->ring_context_gpa;
210
211 if (IS_RESTORE_INHIBIT(shadow_ring_context->ctx_ctrl.val) || skip)
212 return 0;
213
214 s->last_ctx[ring_id].valid = false;
215 context_page_num = workload->engine->context_size;
216 context_page_num = context_page_num >> PAGE_SHIFT;
217
218 if (IS_BROADWELL(gvt->gt->i915) && workload->engine->id == RCS0)
219 context_page_num = 19;
220
221 /* find consecutive GPAs from gma until the first inconsecutive GPA.
222 * read from the continuous GPAs into dst virtual address
223 */
224 gpa_size = 0;
225 for (i = 2; i < context_page_num; i++) {
226 context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
227 (u32)((workload->ctx_desc.lrca + i) <<
228 I915_GTT_PAGE_SHIFT));
229 if (context_gpa == INTEL_GVT_INVALID_ADDR) {
230 gvt_vgpu_err("Invalid guest context descriptor\n");
231 return -EFAULT;
232 }
233
234 if (gpa_size == 0) {
235 gpa_base = context_gpa;
236 dst = context_base + (i << I915_GTT_PAGE_SHIFT);
237 } else if (context_gpa != gpa_base + gpa_size)
238 goto read;
239
240 gpa_size += I915_GTT_PAGE_SIZE;
241
242 if (i == context_page_num - 1)
243 goto read;
244
245 continue;
246
247 read:
248 intel_gvt_read_gpa(vgpu, gpa_base, dst, gpa_size);
249 gpa_base = context_gpa;
250 gpa_size = I915_GTT_PAGE_SIZE;
251 dst = context_base + (i << I915_GTT_PAGE_SHIFT);
252 }
253 ret = intel_gvt_scan_engine_context(workload);
254 if (ret) {
255 gvt_vgpu_err("invalid cmd found in guest context pages\n");
256 return ret;
257 }
258 s->last_ctx[ring_id].valid = true;
259 return 0;
260 }
261
is_gvt_request(struct i915_request * rq)262 static inline bool is_gvt_request(struct i915_request *rq)
263 {
264 return intel_context_force_single_submission(rq->context);
265 }
266
save_ring_hw_state(struct intel_vgpu * vgpu,const struct intel_engine_cs * engine)267 static void save_ring_hw_state(struct intel_vgpu *vgpu,
268 const struct intel_engine_cs *engine)
269 {
270 struct intel_uncore *uncore = engine->uncore;
271 i915_reg_t reg;
272
273 reg = RING_INSTDONE(engine->mmio_base);
274 vgpu_vreg(vgpu, i915_mmio_reg_offset(reg)) =
275 intel_uncore_read(uncore, reg);
276
277 reg = RING_ACTHD(engine->mmio_base);
278 vgpu_vreg(vgpu, i915_mmio_reg_offset(reg)) =
279 intel_uncore_read(uncore, reg);
280
281 reg = RING_ACTHD_UDW(engine->mmio_base);
282 vgpu_vreg(vgpu, i915_mmio_reg_offset(reg)) =
283 intel_uncore_read(uncore, reg);
284 }
285
shadow_context_status_change(struct notifier_block * nb,unsigned long action,void * data)286 static int shadow_context_status_change(struct notifier_block *nb,
287 unsigned long action, void *data)
288 {
289 struct i915_request *rq = data;
290 struct intel_gvt *gvt = container_of(nb, struct intel_gvt,
291 shadow_ctx_notifier_block[rq->engine->id]);
292 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
293 enum intel_engine_id ring_id = rq->engine->id;
294 struct intel_vgpu_workload *workload;
295 unsigned long flags;
296
297 if (!is_gvt_request(rq)) {
298 spin_lock_irqsave(&scheduler->mmio_context_lock, flags);
299 if (action == INTEL_CONTEXT_SCHEDULE_IN &&
300 scheduler->engine_owner[ring_id]) {
301 /* Switch ring from vGPU to host. */
302 intel_gvt_switch_mmio(scheduler->engine_owner[ring_id],
303 NULL, rq->engine);
304 scheduler->engine_owner[ring_id] = NULL;
305 }
306 spin_unlock_irqrestore(&scheduler->mmio_context_lock, flags);
307
308 return NOTIFY_OK;
309 }
310
311 workload = scheduler->current_workload[ring_id];
312 if (unlikely(!workload))
313 return NOTIFY_OK;
314
315 switch (action) {
316 case INTEL_CONTEXT_SCHEDULE_IN:
317 spin_lock_irqsave(&scheduler->mmio_context_lock, flags);
318 if (workload->vgpu != scheduler->engine_owner[ring_id]) {
319 /* Switch ring from host to vGPU or vGPU to vGPU. */
320 intel_gvt_switch_mmio(scheduler->engine_owner[ring_id],
321 workload->vgpu, rq->engine);
322 scheduler->engine_owner[ring_id] = workload->vgpu;
323 } else
324 gvt_dbg_sched("skip ring %d mmio switch for vgpu%d\n",
325 ring_id, workload->vgpu->id);
326 spin_unlock_irqrestore(&scheduler->mmio_context_lock, flags);
327 atomic_set(&workload->shadow_ctx_active, 1);
328 break;
329 case INTEL_CONTEXT_SCHEDULE_OUT:
330 save_ring_hw_state(workload->vgpu, rq->engine);
331 atomic_set(&workload->shadow_ctx_active, 0);
332 break;
333 case INTEL_CONTEXT_SCHEDULE_PREEMPTED:
334 save_ring_hw_state(workload->vgpu, rq->engine);
335 break;
336 default:
337 WARN_ON(1);
338 return NOTIFY_OK;
339 }
340 wake_up(&workload->shadow_ctx_status_wq);
341 return NOTIFY_OK;
342 }
343
344 static void
shadow_context_descriptor_update(struct intel_context * ce,struct intel_vgpu_workload * workload)345 shadow_context_descriptor_update(struct intel_context *ce,
346 struct intel_vgpu_workload *workload)
347 {
348 u64 desc = ce->lrc.desc;
349
350 /*
351 * Update bits 0-11 of the context descriptor which includes flags
352 * like GEN8_CTX_* cached in desc_template
353 */
354 desc &= ~(0x3ull << GEN8_CTX_ADDRESSING_MODE_SHIFT);
355 desc |= (u64)workload->ctx_desc.addressing_mode <<
356 GEN8_CTX_ADDRESSING_MODE_SHIFT;
357
358 ce->lrc.desc = desc;
359 }
360
copy_workload_to_ring_buffer(struct intel_vgpu_workload * workload)361 static int copy_workload_to_ring_buffer(struct intel_vgpu_workload *workload)
362 {
363 struct intel_vgpu *vgpu = workload->vgpu;
364 struct i915_request *req = workload->req;
365 void *shadow_ring_buffer_va;
366 u32 *cs;
367 int err;
368
369 if (GRAPHICS_VER(req->engine->i915) == 9 && is_inhibit_context(req->context))
370 intel_vgpu_restore_inhibit_context(vgpu, req);
371
372 /*
373 * To track whether a request has started on HW, we can emit a
374 * breadcrumb at the beginning of the request and check its
375 * timeline's HWSP to see if the breadcrumb has advanced past the
376 * start of this request. Actually, the request must have the
377 * init_breadcrumb if its timeline set has_init_bread_crumb, or the
378 * scheduler might get a wrong state of it during reset. Since the
379 * requests from gvt always set the has_init_breadcrumb flag, here
380 * need to do the emit_init_breadcrumb for all the requests.
381 */
382 if (req->engine->emit_init_breadcrumb) {
383 err = req->engine->emit_init_breadcrumb(req);
384 if (err) {
385 gvt_vgpu_err("fail to emit init breadcrumb\n");
386 return err;
387 }
388 }
389
390 /* allocate shadow ring buffer */
391 cs = intel_ring_begin(workload->req, workload->rb_len / sizeof(u32));
392 if (IS_ERR(cs)) {
393 gvt_vgpu_err("fail to alloc size =%ld shadow ring buffer\n",
394 workload->rb_len);
395 return PTR_ERR(cs);
396 }
397
398 shadow_ring_buffer_va = workload->shadow_ring_buffer_va;
399
400 /* get shadow ring buffer va */
401 workload->shadow_ring_buffer_va = cs;
402
403 memcpy(cs, shadow_ring_buffer_va,
404 workload->rb_len);
405
406 cs += workload->rb_len / sizeof(u32);
407 intel_ring_advance(workload->req, cs);
408
409 return 0;
410 }
411
release_shadow_wa_ctx(struct intel_shadow_wa_ctx * wa_ctx)412 static void release_shadow_wa_ctx(struct intel_shadow_wa_ctx *wa_ctx)
413 {
414 if (!wa_ctx->indirect_ctx.obj)
415 return;
416
417 i915_gem_object_lock(wa_ctx->indirect_ctx.obj, NULL);
418 i915_gem_object_unpin_map(wa_ctx->indirect_ctx.obj);
419 i915_gem_object_unlock(wa_ctx->indirect_ctx.obj);
420 i915_gem_object_put(wa_ctx->indirect_ctx.obj);
421
422 wa_ctx->indirect_ctx.obj = NULL;
423 wa_ctx->indirect_ctx.shadow_va = NULL;
424 }
425
set_dma_address(struct i915_page_directory * pd,dma_addr_t addr)426 static void set_dma_address(struct i915_page_directory *pd, dma_addr_t addr)
427 {
428 struct scatterlist *sg = pd->pt.base->mm.pages->sgl;
429
430 /* This is not a good idea */
431 sg->dma_address = addr;
432 }
433
set_context_ppgtt_from_shadow(struct intel_vgpu_workload * workload,struct intel_context * ce)434 static void set_context_ppgtt_from_shadow(struct intel_vgpu_workload *workload,
435 struct intel_context *ce)
436 {
437 struct intel_vgpu_mm *mm = workload->shadow_mm;
438 struct i915_ppgtt *ppgtt = i915_vm_to_ppgtt(ce->vm);
439 int i = 0;
440
441 if (mm->ppgtt_mm.root_entry_type == GTT_TYPE_PPGTT_ROOT_L4_ENTRY) {
442 set_dma_address(ppgtt->pd, mm->ppgtt_mm.shadow_pdps[0]);
443 } else {
444 for (i = 0; i < GVT_RING_CTX_NR_PDPS; i++) {
445 struct i915_page_directory * const pd =
446 i915_pd_entry(ppgtt->pd, i);
447 /* skip now as current i915 ppgtt alloc won't allocate
448 top level pdp for non 4-level table, won't impact
449 shadow ppgtt. */
450 if (!pd)
451 break;
452
453 set_dma_address(pd, mm->ppgtt_mm.shadow_pdps[i]);
454 }
455 }
456 }
457
458 static int
intel_gvt_workload_req_alloc(struct intel_vgpu_workload * workload)459 intel_gvt_workload_req_alloc(struct intel_vgpu_workload *workload)
460 {
461 struct intel_vgpu *vgpu = workload->vgpu;
462 struct intel_vgpu_submission *s = &vgpu->submission;
463 struct i915_request *rq;
464
465 if (workload->req)
466 return 0;
467
468 rq = i915_request_create(s->shadow[workload->engine->id]);
469 if (IS_ERR(rq)) {
470 gvt_vgpu_err("fail to allocate gem request\n");
471 return PTR_ERR(rq);
472 }
473
474 workload->req = i915_request_get(rq);
475 return 0;
476 }
477
478 /**
479 * intel_gvt_scan_and_shadow_workload - audit the workload by scanning and
480 * shadow it as well, include ringbuffer,wa_ctx and ctx.
481 * @workload: an abstract entity for each execlist submission.
482 *
483 * This function is called before the workload submitting to i915, to make
484 * sure the content of the workload is valid.
485 */
intel_gvt_scan_and_shadow_workload(struct intel_vgpu_workload * workload)486 int intel_gvt_scan_and_shadow_workload(struct intel_vgpu_workload *workload)
487 {
488 struct intel_vgpu *vgpu = workload->vgpu;
489 struct intel_vgpu_submission *s = &vgpu->submission;
490 int ret;
491
492 lockdep_assert_held(&vgpu->vgpu_lock);
493
494 if (workload->shadow)
495 return 0;
496
497 if (!test_and_set_bit(workload->engine->id, s->shadow_ctx_desc_updated))
498 shadow_context_descriptor_update(s->shadow[workload->engine->id],
499 workload);
500
501 ret = intel_gvt_scan_and_shadow_ringbuffer(workload);
502 if (ret)
503 return ret;
504
505 if (workload->engine->id == RCS0 &&
506 workload->wa_ctx.indirect_ctx.size) {
507 ret = intel_gvt_scan_and_shadow_wa_ctx(&workload->wa_ctx);
508 if (ret)
509 goto err_shadow;
510 }
511
512 workload->shadow = true;
513 return 0;
514
515 err_shadow:
516 release_shadow_wa_ctx(&workload->wa_ctx);
517 return ret;
518 }
519
520 static void release_shadow_batch_buffer(struct intel_vgpu_workload *workload);
521
prepare_shadow_batch_buffer(struct intel_vgpu_workload * workload)522 static int prepare_shadow_batch_buffer(struct intel_vgpu_workload *workload)
523 {
524 struct intel_gvt *gvt = workload->vgpu->gvt;
525 const int gmadr_bytes = gvt->device_info.gmadr_bytes_in_cmd;
526 struct intel_vgpu_shadow_bb *bb;
527 struct i915_gem_ww_ctx ww;
528 int ret;
529
530 list_for_each_entry(bb, &workload->shadow_bb, list) {
531 /*
532 * For privilege batch buffer and not wa_ctx, the bb_start_cmd_va
533 * is only updated into ring_scan_buffer, not real ring address
534 * allocated in later copy_workload_to_ring_buffer. Please be noted
535 * shadow_ring_buffer_va is now pointed to real ring buffer va
536 * in copy_workload_to_ring_buffer.
537 */
538
539 if (bb->bb_offset)
540 bb->bb_start_cmd_va = workload->shadow_ring_buffer_va
541 + bb->bb_offset;
542
543 /*
544 * For non-priv bb, scan&shadow is only for
545 * debugging purpose, so the content of shadow bb
546 * is the same as original bb. Therefore,
547 * here, rather than switch to shadow bb's gma
548 * address, we directly use original batch buffer's
549 * gma address, and send original bb to hardware
550 * directly.
551 */
552 if (!bb->ppgtt) {
553 i915_gem_ww_ctx_init(&ww, false);
554 retry:
555 i915_gem_object_lock(bb->obj, &ww);
556
557 bb->vma = i915_gem_object_ggtt_pin_ww(bb->obj, &ww,
558 NULL, 0, 0, 0);
559 if (IS_ERR(bb->vma)) {
560 ret = PTR_ERR(bb->vma);
561 if (ret == -EDEADLK) {
562 ret = i915_gem_ww_ctx_backoff(&ww);
563 if (!ret)
564 goto retry;
565 }
566 goto err;
567 }
568
569 /* relocate shadow batch buffer */
570 bb->bb_start_cmd_va[1] = i915_ggtt_offset(bb->vma);
571 if (gmadr_bytes == 8)
572 bb->bb_start_cmd_va[2] = 0;
573
574 ret = i915_vma_move_to_active(bb->vma, workload->req,
575 __EXEC_OBJECT_NO_REQUEST_AWAIT);
576 if (ret)
577 goto err;
578
579 /* No one is going to touch shadow bb from now on. */
580 i915_gem_object_flush_map(bb->obj);
581 i915_gem_ww_ctx_fini(&ww);
582 }
583 }
584 return 0;
585 err:
586 i915_gem_ww_ctx_fini(&ww);
587 release_shadow_batch_buffer(workload);
588 return ret;
589 }
590
update_wa_ctx_2_shadow_ctx(struct intel_shadow_wa_ctx * wa_ctx)591 static void update_wa_ctx_2_shadow_ctx(struct intel_shadow_wa_ctx *wa_ctx)
592 {
593 struct intel_vgpu_workload *workload =
594 container_of(wa_ctx, struct intel_vgpu_workload, wa_ctx);
595 struct i915_request *rq = workload->req;
596 struct execlist_ring_context *shadow_ring_context =
597 (struct execlist_ring_context *)rq->context->lrc_reg_state;
598
599 shadow_ring_context->bb_per_ctx_ptr.val =
600 (shadow_ring_context->bb_per_ctx_ptr.val &
601 (~PER_CTX_ADDR_MASK)) | wa_ctx->per_ctx.shadow_gma;
602 shadow_ring_context->rcs_indirect_ctx.val =
603 (shadow_ring_context->rcs_indirect_ctx.val &
604 (~INDIRECT_CTX_ADDR_MASK)) | wa_ctx->indirect_ctx.shadow_gma;
605 }
606
prepare_shadow_wa_ctx(struct intel_shadow_wa_ctx * wa_ctx)607 static int prepare_shadow_wa_ctx(struct intel_shadow_wa_ctx *wa_ctx)
608 {
609 struct i915_vma *vma;
610 unsigned char *per_ctx_va =
611 (unsigned char *)wa_ctx->indirect_ctx.shadow_va +
612 wa_ctx->indirect_ctx.size;
613 struct i915_gem_ww_ctx ww;
614 int ret;
615
616 if (wa_ctx->indirect_ctx.size == 0)
617 return 0;
618
619 i915_gem_ww_ctx_init(&ww, false);
620 retry:
621 i915_gem_object_lock(wa_ctx->indirect_ctx.obj, &ww);
622
623 vma = i915_gem_object_ggtt_pin_ww(wa_ctx->indirect_ctx.obj, &ww, NULL,
624 0, CACHELINE_BYTES, 0);
625 if (IS_ERR(vma)) {
626 ret = PTR_ERR(vma);
627 if (ret == -EDEADLK) {
628 ret = i915_gem_ww_ctx_backoff(&ww);
629 if (!ret)
630 goto retry;
631 }
632 return ret;
633 }
634
635 i915_gem_ww_ctx_fini(&ww);
636
637 /* FIXME: we are not tracking our pinned VMA leaving it
638 * up to the core to fix up the stray pin_count upon
639 * free.
640 */
641
642 wa_ctx->indirect_ctx.shadow_gma = i915_ggtt_offset(vma);
643
644 wa_ctx->per_ctx.shadow_gma = *((unsigned int *)per_ctx_va + 1);
645 memset(per_ctx_va, 0, CACHELINE_BYTES);
646
647 update_wa_ctx_2_shadow_ctx(wa_ctx);
648 return 0;
649 }
650
update_vreg_in_ctx(struct intel_vgpu_workload * workload)651 static void update_vreg_in_ctx(struct intel_vgpu_workload *workload)
652 {
653 vgpu_vreg_t(workload->vgpu, RING_START(workload->engine->mmio_base)) =
654 workload->rb_start;
655 }
656
release_shadow_batch_buffer(struct intel_vgpu_workload * workload)657 static void release_shadow_batch_buffer(struct intel_vgpu_workload *workload)
658 {
659 struct intel_vgpu_shadow_bb *bb, *pos;
660
661 if (list_empty(&workload->shadow_bb))
662 return;
663
664 bb = list_first_entry(&workload->shadow_bb,
665 struct intel_vgpu_shadow_bb, list);
666
667 list_for_each_entry_safe(bb, pos, &workload->shadow_bb, list) {
668 if (bb->obj) {
669 i915_gem_object_lock(bb->obj, NULL);
670 if (bb->va && !IS_ERR(bb->va))
671 i915_gem_object_unpin_map(bb->obj);
672
673 if (bb->vma && !IS_ERR(bb->vma))
674 i915_vma_unpin(bb->vma);
675
676 i915_gem_object_unlock(bb->obj);
677 i915_gem_object_put(bb->obj);
678 }
679 list_del(&bb->list);
680 kfree(bb);
681 }
682 }
683
684 static int
intel_vgpu_shadow_mm_pin(struct intel_vgpu_workload * workload)685 intel_vgpu_shadow_mm_pin(struct intel_vgpu_workload *workload)
686 {
687 struct intel_vgpu *vgpu = workload->vgpu;
688 struct intel_vgpu_mm *m;
689 int ret = 0;
690
691 ret = intel_vgpu_pin_mm(workload->shadow_mm);
692 if (ret) {
693 gvt_vgpu_err("fail to vgpu pin mm\n");
694 return ret;
695 }
696
697 if (workload->shadow_mm->type != INTEL_GVT_MM_PPGTT ||
698 !workload->shadow_mm->ppgtt_mm.shadowed) {
699 intel_vgpu_unpin_mm(workload->shadow_mm);
700 gvt_vgpu_err("workload shadow ppgtt isn't ready\n");
701 return -EINVAL;
702 }
703
704 if (!list_empty(&workload->lri_shadow_mm)) {
705 list_for_each_entry(m, &workload->lri_shadow_mm,
706 ppgtt_mm.link) {
707 ret = intel_vgpu_pin_mm(m);
708 if (ret) {
709 list_for_each_entry_from_reverse(m,
710 &workload->lri_shadow_mm,
711 ppgtt_mm.link)
712 intel_vgpu_unpin_mm(m);
713 gvt_vgpu_err("LRI shadow ppgtt fail to pin\n");
714 break;
715 }
716 }
717 }
718
719 if (ret)
720 intel_vgpu_unpin_mm(workload->shadow_mm);
721
722 return ret;
723 }
724
725 static void
intel_vgpu_shadow_mm_unpin(struct intel_vgpu_workload * workload)726 intel_vgpu_shadow_mm_unpin(struct intel_vgpu_workload *workload)
727 {
728 struct intel_vgpu_mm *m;
729
730 if (!list_empty(&workload->lri_shadow_mm)) {
731 list_for_each_entry(m, &workload->lri_shadow_mm,
732 ppgtt_mm.link)
733 intel_vgpu_unpin_mm(m);
734 }
735 intel_vgpu_unpin_mm(workload->shadow_mm);
736 }
737
prepare_workload(struct intel_vgpu_workload * workload)738 static int prepare_workload(struct intel_vgpu_workload *workload)
739 {
740 struct intel_vgpu *vgpu = workload->vgpu;
741 struct intel_vgpu_submission *s = &vgpu->submission;
742 int ret = 0;
743
744 ret = intel_vgpu_shadow_mm_pin(workload);
745 if (ret) {
746 gvt_vgpu_err("fail to pin shadow mm\n");
747 return ret;
748 }
749
750 update_shadow_pdps(workload);
751
752 set_context_ppgtt_from_shadow(workload, s->shadow[workload->engine->id]);
753
754 ret = intel_vgpu_sync_oos_pages(workload->vgpu);
755 if (ret) {
756 gvt_vgpu_err("fail to vgpu sync oos pages\n");
757 goto err_unpin_mm;
758 }
759
760 ret = intel_vgpu_flush_post_shadow(workload->vgpu);
761 if (ret) {
762 gvt_vgpu_err("fail to flush post shadow\n");
763 goto err_unpin_mm;
764 }
765
766 ret = copy_workload_to_ring_buffer(workload);
767 if (ret) {
768 gvt_vgpu_err("fail to generate request\n");
769 goto err_unpin_mm;
770 }
771
772 ret = prepare_shadow_batch_buffer(workload);
773 if (ret) {
774 gvt_vgpu_err("fail to prepare_shadow_batch_buffer\n");
775 goto err_unpin_mm;
776 }
777
778 ret = prepare_shadow_wa_ctx(&workload->wa_ctx);
779 if (ret) {
780 gvt_vgpu_err("fail to prepare_shadow_wa_ctx\n");
781 goto err_shadow_batch;
782 }
783
784 if (workload->prepare) {
785 ret = workload->prepare(workload);
786 if (ret)
787 goto err_shadow_wa_ctx;
788 }
789
790 return 0;
791 err_shadow_wa_ctx:
792 release_shadow_wa_ctx(&workload->wa_ctx);
793 err_shadow_batch:
794 release_shadow_batch_buffer(workload);
795 err_unpin_mm:
796 intel_vgpu_shadow_mm_unpin(workload);
797 return ret;
798 }
799
dispatch_workload(struct intel_vgpu_workload * workload)800 static int dispatch_workload(struct intel_vgpu_workload *workload)
801 {
802 struct intel_vgpu *vgpu = workload->vgpu;
803 struct i915_request *rq;
804 int ret;
805
806 gvt_dbg_sched("ring id %s prepare to dispatch workload %p\n",
807 workload->engine->name, workload);
808
809 mutex_lock(&vgpu->vgpu_lock);
810
811 ret = intel_gvt_workload_req_alloc(workload);
812 if (ret)
813 goto err_req;
814
815 ret = intel_gvt_scan_and_shadow_workload(workload);
816 if (ret)
817 goto out;
818
819 ret = populate_shadow_context(workload);
820 if (ret) {
821 release_shadow_wa_ctx(&workload->wa_ctx);
822 goto out;
823 }
824
825 ret = prepare_workload(workload);
826 out:
827 if (ret) {
828 /* We might still need to add request with
829 * clean ctx to retire it properly..
830 */
831 rq = fetch_and_zero(&workload->req);
832 i915_request_put(rq);
833 }
834
835 if (!IS_ERR_OR_NULL(workload->req)) {
836 gvt_dbg_sched("ring id %s submit workload to i915 %p\n",
837 workload->engine->name, workload->req);
838 i915_request_add(workload->req);
839 workload->dispatched = true;
840 }
841 err_req:
842 if (ret)
843 workload->status = ret;
844 mutex_unlock(&vgpu->vgpu_lock);
845 return ret;
846 }
847
848 static struct intel_vgpu_workload *
pick_next_workload(struct intel_gvt * gvt,struct intel_engine_cs * engine)849 pick_next_workload(struct intel_gvt *gvt, struct intel_engine_cs *engine)
850 {
851 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
852 struct intel_vgpu_workload *workload = NULL;
853
854 mutex_lock(&gvt->sched_lock);
855
856 /*
857 * no current vgpu / will be scheduled out / no workload
858 * bail out
859 */
860 if (!scheduler->current_vgpu) {
861 gvt_dbg_sched("ring %s stop - no current vgpu\n", engine->name);
862 goto out;
863 }
864
865 if (scheduler->need_reschedule) {
866 gvt_dbg_sched("ring %s stop - will reschedule\n", engine->name);
867 goto out;
868 }
869
870 if (!test_bit(INTEL_VGPU_STATUS_ACTIVE,
871 scheduler->current_vgpu->status) ||
872 list_empty(workload_q_head(scheduler->current_vgpu, engine)))
873 goto out;
874
875 /*
876 * still have current workload, maybe the workload disptacher
877 * fail to submit it for some reason, resubmit it.
878 */
879 if (scheduler->current_workload[engine->id]) {
880 workload = scheduler->current_workload[engine->id];
881 gvt_dbg_sched("ring %s still have current workload %p\n",
882 engine->name, workload);
883 goto out;
884 }
885
886 /*
887 * pick a workload as current workload
888 * once current workload is set, schedule policy routines
889 * will wait the current workload is finished when trying to
890 * schedule out a vgpu.
891 */
892 scheduler->current_workload[engine->id] =
893 list_first_entry(workload_q_head(scheduler->current_vgpu,
894 engine),
895 struct intel_vgpu_workload, list);
896
897 workload = scheduler->current_workload[engine->id];
898
899 gvt_dbg_sched("ring %s pick new workload %p\n", engine->name, workload);
900
901 atomic_inc(&workload->vgpu->submission.running_workload_num);
902 out:
903 mutex_unlock(&gvt->sched_lock);
904 return workload;
905 }
906
update_guest_pdps(struct intel_vgpu * vgpu,u64 ring_context_gpa,u32 pdp[8])907 static void update_guest_pdps(struct intel_vgpu *vgpu,
908 u64 ring_context_gpa, u32 pdp[8])
909 {
910 u64 gpa;
911 int i;
912
913 gpa = ring_context_gpa + RING_CTX_OFF(pdps[0].val);
914
915 for (i = 0; i < 8; i++)
916 intel_gvt_write_gpa(vgpu, gpa + i * 8, &pdp[7 - i], 4);
917 }
918
919 static __maybe_unused bool
check_shadow_context_ppgtt(struct execlist_ring_context * c,struct intel_vgpu_mm * m)920 check_shadow_context_ppgtt(struct execlist_ring_context *c, struct intel_vgpu_mm *m)
921 {
922 if (m->ppgtt_mm.root_entry_type == GTT_TYPE_PPGTT_ROOT_L4_ENTRY) {
923 u64 shadow_pdp = c->pdps[7].val | (u64) c->pdps[6].val << 32;
924
925 if (shadow_pdp != m->ppgtt_mm.shadow_pdps[0]) {
926 gvt_dbg_mm("4-level context ppgtt not match LRI command\n");
927 return false;
928 }
929 return true;
930 } else {
931 /* see comment in LRI handler in cmd_parser.c */
932 gvt_dbg_mm("invalid shadow mm type\n");
933 return false;
934 }
935 }
936
update_guest_context(struct intel_vgpu_workload * workload)937 static void update_guest_context(struct intel_vgpu_workload *workload)
938 {
939 struct i915_request *rq = workload->req;
940 struct intel_vgpu *vgpu = workload->vgpu;
941 struct execlist_ring_context *shadow_ring_context;
942 struct intel_context *ctx = workload->req->context;
943 void *context_base;
944 void *src;
945 unsigned long context_gpa, context_page_num;
946 unsigned long gpa_base; /* first gpa of consecutive GPAs */
947 unsigned long gpa_size; /* size of consecutive GPAs*/
948 int i;
949 u32 ring_base;
950 u32 head, tail;
951 u16 wrap_count;
952
953 gvt_dbg_sched("ring id %d workload lrca %x\n", rq->engine->id,
954 workload->ctx_desc.lrca);
955
956 GEM_BUG_ON(!intel_context_is_pinned(ctx));
957
958 head = workload->rb_head;
959 tail = workload->rb_tail;
960 wrap_count = workload->guest_rb_head >> RB_HEAD_WRAP_CNT_OFF;
961
962 if (tail < head) {
963 if (wrap_count == RB_HEAD_WRAP_CNT_MAX)
964 wrap_count = 0;
965 else
966 wrap_count += 1;
967 }
968
969 head = (wrap_count << RB_HEAD_WRAP_CNT_OFF) | tail;
970
971 ring_base = rq->engine->mmio_base;
972 vgpu_vreg_t(vgpu, RING_TAIL(ring_base)) = tail;
973 vgpu_vreg_t(vgpu, RING_HEAD(ring_base)) = head;
974
975 context_page_num = rq->engine->context_size;
976 context_page_num = context_page_num >> PAGE_SHIFT;
977
978 if (IS_BROADWELL(rq->i915) && rq->engine->id == RCS0)
979 context_page_num = 19;
980
981 context_base = (void *) ctx->lrc_reg_state -
982 (LRC_STATE_PN << I915_GTT_PAGE_SHIFT);
983
984 /* find consecutive GPAs from gma until the first inconsecutive GPA.
985 * write to the consecutive GPAs from src virtual address
986 */
987 gpa_size = 0;
988 for (i = 2; i < context_page_num; i++) {
989 context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
990 (u32)((workload->ctx_desc.lrca + i) <<
991 I915_GTT_PAGE_SHIFT));
992 if (context_gpa == INTEL_GVT_INVALID_ADDR) {
993 gvt_vgpu_err("invalid guest context descriptor\n");
994 return;
995 }
996
997 if (gpa_size == 0) {
998 gpa_base = context_gpa;
999 src = context_base + (i << I915_GTT_PAGE_SHIFT);
1000 } else if (context_gpa != gpa_base + gpa_size)
1001 goto write;
1002
1003 gpa_size += I915_GTT_PAGE_SIZE;
1004
1005 if (i == context_page_num - 1)
1006 goto write;
1007
1008 continue;
1009
1010 write:
1011 intel_gvt_write_gpa(vgpu, gpa_base, src, gpa_size);
1012 gpa_base = context_gpa;
1013 gpa_size = I915_GTT_PAGE_SIZE;
1014 src = context_base + (i << I915_GTT_PAGE_SHIFT);
1015 }
1016
1017 intel_gvt_write_gpa(vgpu, workload->ring_context_gpa +
1018 RING_CTX_OFF(ring_header.val), &workload->rb_tail, 4);
1019
1020 shadow_ring_context = (void *) ctx->lrc_reg_state;
1021
1022 if (!list_empty(&workload->lri_shadow_mm)) {
1023 struct intel_vgpu_mm *m = list_last_entry(&workload->lri_shadow_mm,
1024 struct intel_vgpu_mm,
1025 ppgtt_mm.link);
1026 GEM_BUG_ON(!check_shadow_context_ppgtt(shadow_ring_context, m));
1027 update_guest_pdps(vgpu, workload->ring_context_gpa,
1028 (void *)m->ppgtt_mm.guest_pdps);
1029 }
1030
1031 #define COPY_REG(name) \
1032 intel_gvt_write_gpa(vgpu, workload->ring_context_gpa + \
1033 RING_CTX_OFF(name.val), &shadow_ring_context->name.val, 4)
1034
1035 COPY_REG(ctx_ctrl);
1036 COPY_REG(ctx_timestamp);
1037
1038 #undef COPY_REG
1039
1040 intel_gvt_write_gpa(vgpu,
1041 workload->ring_context_gpa +
1042 sizeof(*shadow_ring_context),
1043 (void *)shadow_ring_context +
1044 sizeof(*shadow_ring_context),
1045 I915_GTT_PAGE_SIZE - sizeof(*shadow_ring_context));
1046 }
1047
intel_vgpu_clean_workloads(struct intel_vgpu * vgpu,intel_engine_mask_t engine_mask)1048 void intel_vgpu_clean_workloads(struct intel_vgpu *vgpu,
1049 intel_engine_mask_t engine_mask)
1050 {
1051 struct intel_vgpu_submission *s = &vgpu->submission;
1052 struct intel_engine_cs *engine;
1053 struct intel_vgpu_workload *pos, *n;
1054 intel_engine_mask_t tmp;
1055
1056 /* free the unsubmitted workloads in the queues. */
1057 for_each_engine_masked(engine, vgpu->gvt->gt, engine_mask, tmp) {
1058 list_for_each_entry_safe(pos, n,
1059 &s->workload_q_head[engine->id], list) {
1060 list_del_init(&pos->list);
1061 intel_vgpu_destroy_workload(pos);
1062 }
1063 clear_bit(engine->id, s->shadow_ctx_desc_updated);
1064 }
1065 }
1066
complete_current_workload(struct intel_gvt * gvt,int ring_id)1067 static void complete_current_workload(struct intel_gvt *gvt, int ring_id)
1068 {
1069 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
1070 struct intel_vgpu_workload *workload =
1071 scheduler->current_workload[ring_id];
1072 struct intel_vgpu *vgpu = workload->vgpu;
1073 struct intel_vgpu_submission *s = &vgpu->submission;
1074 struct i915_request *rq = workload->req;
1075 int event;
1076
1077 mutex_lock(&vgpu->vgpu_lock);
1078 mutex_lock(&gvt->sched_lock);
1079
1080 /* For the workload w/ request, needs to wait for the context
1081 * switch to make sure request is completed.
1082 * For the workload w/o request, directly complete the workload.
1083 */
1084 if (rq) {
1085 wait_event(workload->shadow_ctx_status_wq,
1086 !atomic_read(&workload->shadow_ctx_active));
1087
1088 /* If this request caused GPU hang, req->fence.error will
1089 * be set to -EIO. Use -EIO to set workload status so
1090 * that when this request caused GPU hang, didn't trigger
1091 * context switch interrupt to guest.
1092 */
1093 if (likely(workload->status == -EINPROGRESS)) {
1094 if (workload->req->fence.error == -EIO)
1095 workload->status = -EIO;
1096 else
1097 workload->status = 0;
1098 }
1099
1100 if (!workload->status &&
1101 !(vgpu->resetting_eng & BIT(ring_id))) {
1102 update_guest_context(workload);
1103
1104 for_each_set_bit(event, workload->pending_events,
1105 INTEL_GVT_EVENT_MAX)
1106 intel_vgpu_trigger_virtual_event(vgpu, event);
1107 }
1108
1109 i915_request_put(fetch_and_zero(&workload->req));
1110 }
1111
1112 gvt_dbg_sched("ring id %d complete workload %p status %d\n",
1113 ring_id, workload, workload->status);
1114
1115 scheduler->current_workload[ring_id] = NULL;
1116
1117 list_del_init(&workload->list);
1118
1119 if (workload->status || vgpu->resetting_eng & BIT(ring_id)) {
1120 /* if workload->status is not successful means HW GPU
1121 * has occurred GPU hang or something wrong with i915/GVT,
1122 * and GVT won't inject context switch interrupt to guest.
1123 * So this error is a vGPU hang actually to the guest.
1124 * According to this we should emunlate a vGPU hang. If
1125 * there are pending workloads which are already submitted
1126 * from guest, we should clean them up like HW GPU does.
1127 *
1128 * if it is in middle of engine resetting, the pending
1129 * workloads won't be submitted to HW GPU and will be
1130 * cleaned up during the resetting process later, so doing
1131 * the workload clean up here doesn't have any impact.
1132 **/
1133 intel_vgpu_clean_workloads(vgpu, BIT(ring_id));
1134 }
1135
1136 workload->complete(workload);
1137
1138 intel_vgpu_shadow_mm_unpin(workload);
1139 intel_vgpu_destroy_workload(workload);
1140
1141 atomic_dec(&s->running_workload_num);
1142 wake_up(&scheduler->workload_complete_wq);
1143
1144 if (gvt->scheduler.need_reschedule)
1145 intel_gvt_request_service(gvt, INTEL_GVT_REQUEST_EVENT_SCHED);
1146
1147 mutex_unlock(&gvt->sched_lock);
1148 mutex_unlock(&vgpu->vgpu_lock);
1149 }
1150
workload_thread(void * arg)1151 static int workload_thread(void *arg)
1152 {
1153 struct intel_engine_cs *engine = arg;
1154 const bool need_force_wake = GRAPHICS_VER(engine->i915) >= 9;
1155 struct intel_gvt *gvt = engine->i915->gvt;
1156 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
1157 struct intel_vgpu_workload *workload = NULL;
1158 struct intel_vgpu *vgpu = NULL;
1159 int ret;
1160 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1161
1162 gvt_dbg_core("workload thread for ring %s started\n", engine->name);
1163
1164 while (!kthread_should_stop()) {
1165 intel_wakeref_t wakeref;
1166
1167 add_wait_queue(&scheduler->waitq[engine->id], &wait);
1168 do {
1169 workload = pick_next_workload(gvt, engine);
1170 if (workload)
1171 break;
1172 wait_woken(&wait, TASK_INTERRUPTIBLE,
1173 MAX_SCHEDULE_TIMEOUT);
1174 } while (!kthread_should_stop());
1175 remove_wait_queue(&scheduler->waitq[engine->id], &wait);
1176
1177 if (!workload)
1178 break;
1179
1180 gvt_dbg_sched("ring %s next workload %p vgpu %d\n",
1181 engine->name, workload,
1182 workload->vgpu->id);
1183
1184 wakeref = intel_runtime_pm_get(engine->uncore->rpm);
1185
1186 gvt_dbg_sched("ring %s will dispatch workload %p\n",
1187 engine->name, workload);
1188
1189 if (need_force_wake)
1190 intel_uncore_forcewake_get(engine->uncore,
1191 FORCEWAKE_ALL);
1192 /*
1193 * Update the vReg of the vGPU which submitted this
1194 * workload. The vGPU may use these registers for checking
1195 * the context state. The value comes from GPU commands
1196 * in this workload.
1197 */
1198 update_vreg_in_ctx(workload);
1199
1200 ret = dispatch_workload(workload);
1201
1202 if (ret) {
1203 vgpu = workload->vgpu;
1204 gvt_vgpu_err("fail to dispatch workload, skip\n");
1205 goto complete;
1206 }
1207
1208 gvt_dbg_sched("ring %s wait workload %p\n",
1209 engine->name, workload);
1210 i915_request_wait(workload->req, 0, MAX_SCHEDULE_TIMEOUT);
1211
1212 complete:
1213 gvt_dbg_sched("will complete workload %p, status: %d\n",
1214 workload, workload->status);
1215
1216 complete_current_workload(gvt, engine->id);
1217
1218 if (need_force_wake)
1219 intel_uncore_forcewake_put(engine->uncore,
1220 FORCEWAKE_ALL);
1221
1222 intel_runtime_pm_put(engine->uncore->rpm, wakeref);
1223 if (ret && (vgpu_is_vm_unhealthy(ret)))
1224 enter_failsafe_mode(vgpu, GVT_FAILSAFE_GUEST_ERR);
1225 }
1226 return 0;
1227 }
1228
intel_gvt_wait_vgpu_idle(struct intel_vgpu * vgpu)1229 void intel_gvt_wait_vgpu_idle(struct intel_vgpu *vgpu)
1230 {
1231 struct intel_vgpu_submission *s = &vgpu->submission;
1232 struct intel_gvt *gvt = vgpu->gvt;
1233 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
1234
1235 if (atomic_read(&s->running_workload_num)) {
1236 gvt_dbg_sched("wait vgpu idle\n");
1237
1238 wait_event(scheduler->workload_complete_wq,
1239 !atomic_read(&s->running_workload_num));
1240 }
1241 }
1242
intel_gvt_clean_workload_scheduler(struct intel_gvt * gvt)1243 void intel_gvt_clean_workload_scheduler(struct intel_gvt *gvt)
1244 {
1245 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
1246 struct intel_engine_cs *engine;
1247 enum intel_engine_id i;
1248
1249 gvt_dbg_core("clean workload scheduler\n");
1250
1251 for_each_engine(engine, gvt->gt, i) {
1252 atomic_notifier_chain_unregister(
1253 &engine->context_status_notifier,
1254 &gvt->shadow_ctx_notifier_block[i]);
1255 kthread_stop(scheduler->thread[i]);
1256 }
1257 }
1258
intel_gvt_init_workload_scheduler(struct intel_gvt * gvt)1259 int intel_gvt_init_workload_scheduler(struct intel_gvt *gvt)
1260 {
1261 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
1262 struct intel_engine_cs *engine;
1263 enum intel_engine_id i;
1264 int ret;
1265
1266 gvt_dbg_core("init workload scheduler\n");
1267
1268 init_waitqueue_head(&scheduler->workload_complete_wq);
1269
1270 for_each_engine(engine, gvt->gt, i) {
1271 init_waitqueue_head(&scheduler->waitq[i]);
1272
1273 scheduler->thread[i] = kthread_run(workload_thread, engine,
1274 "gvt:%s", engine->name);
1275 if (IS_ERR(scheduler->thread[i])) {
1276 gvt_err("fail to create workload thread\n");
1277 ret = PTR_ERR(scheduler->thread[i]);
1278 goto err;
1279 }
1280
1281 gvt->shadow_ctx_notifier_block[i].notifier_call =
1282 shadow_context_status_change;
1283 atomic_notifier_chain_register(&engine->context_status_notifier,
1284 &gvt->shadow_ctx_notifier_block[i]);
1285 }
1286
1287 return 0;
1288
1289 err:
1290 intel_gvt_clean_workload_scheduler(gvt);
1291 return ret;
1292 }
1293
1294 static void
i915_context_ppgtt_root_restore(struct intel_vgpu_submission * s,struct i915_ppgtt * ppgtt)1295 i915_context_ppgtt_root_restore(struct intel_vgpu_submission *s,
1296 struct i915_ppgtt *ppgtt)
1297 {
1298 int i;
1299
1300 if (i915_vm_is_4lvl(&ppgtt->vm)) {
1301 set_dma_address(ppgtt->pd, s->i915_context_pml4);
1302 } else {
1303 for (i = 0; i < GEN8_3LVL_PDPES; i++) {
1304 struct i915_page_directory * const pd =
1305 i915_pd_entry(ppgtt->pd, i);
1306
1307 set_dma_address(pd, s->i915_context_pdps[i]);
1308 }
1309 }
1310 }
1311
1312 /**
1313 * intel_vgpu_clean_submission - free submission-related resource for vGPU
1314 * @vgpu: a vGPU
1315 *
1316 * This function is called when a vGPU is being destroyed.
1317 *
1318 */
intel_vgpu_clean_submission(struct intel_vgpu * vgpu)1319 void intel_vgpu_clean_submission(struct intel_vgpu *vgpu)
1320 {
1321 struct intel_vgpu_submission *s = &vgpu->submission;
1322 struct intel_engine_cs *engine;
1323 enum intel_engine_id id;
1324
1325 intel_vgpu_select_submission_ops(vgpu, ALL_ENGINES, 0);
1326
1327 i915_context_ppgtt_root_restore(s, i915_vm_to_ppgtt(s->shadow[0]->vm));
1328 for_each_engine(engine, vgpu->gvt->gt, id)
1329 intel_context_put(s->shadow[id]);
1330
1331 kmem_cache_destroy(s->workloads);
1332 }
1333
1334
1335 /**
1336 * intel_vgpu_reset_submission - reset submission-related resource for vGPU
1337 * @vgpu: a vGPU
1338 * @engine_mask: engines expected to be reset
1339 *
1340 * This function is called when a vGPU is being destroyed.
1341 *
1342 */
intel_vgpu_reset_submission(struct intel_vgpu * vgpu,intel_engine_mask_t engine_mask)1343 void intel_vgpu_reset_submission(struct intel_vgpu *vgpu,
1344 intel_engine_mask_t engine_mask)
1345 {
1346 struct intel_vgpu_submission *s = &vgpu->submission;
1347
1348 if (!s->active)
1349 return;
1350
1351 intel_vgpu_clean_workloads(vgpu, engine_mask);
1352 s->ops->reset(vgpu, engine_mask);
1353 }
1354
1355 static void
i915_context_ppgtt_root_save(struct intel_vgpu_submission * s,struct i915_ppgtt * ppgtt)1356 i915_context_ppgtt_root_save(struct intel_vgpu_submission *s,
1357 struct i915_ppgtt *ppgtt)
1358 {
1359 int i;
1360
1361 if (i915_vm_is_4lvl(&ppgtt->vm)) {
1362 s->i915_context_pml4 = px_dma(ppgtt->pd);
1363 } else {
1364 for (i = 0; i < GEN8_3LVL_PDPES; i++) {
1365 struct i915_page_directory * const pd =
1366 i915_pd_entry(ppgtt->pd, i);
1367
1368 s->i915_context_pdps[i] = px_dma(pd);
1369 }
1370 }
1371 }
1372
1373 /**
1374 * intel_vgpu_setup_submission - setup submission-related resource for vGPU
1375 * @vgpu: a vGPU
1376 *
1377 * This function is called when a vGPU is being created.
1378 *
1379 * Returns:
1380 * Zero on success, negative error code if failed.
1381 *
1382 */
intel_vgpu_setup_submission(struct intel_vgpu * vgpu)1383 int intel_vgpu_setup_submission(struct intel_vgpu *vgpu)
1384 {
1385 struct drm_i915_private *i915 = vgpu->gvt->gt->i915;
1386 struct intel_vgpu_submission *s = &vgpu->submission;
1387 struct intel_engine_cs *engine;
1388 struct i915_ppgtt *ppgtt;
1389 enum intel_engine_id i;
1390 int ret;
1391
1392 ppgtt = i915_ppgtt_create(to_gt(i915), I915_BO_ALLOC_PM_EARLY);
1393 if (IS_ERR(ppgtt))
1394 return PTR_ERR(ppgtt);
1395
1396 i915_context_ppgtt_root_save(s, ppgtt);
1397
1398 for_each_engine(engine, vgpu->gvt->gt, i) {
1399 struct intel_context *ce;
1400
1401 INIT_LIST_HEAD(&s->workload_q_head[i]);
1402 s->shadow[i] = ERR_PTR(-EINVAL);
1403
1404 ce = intel_context_create(engine);
1405 if (IS_ERR(ce)) {
1406 ret = PTR_ERR(ce);
1407 goto out_shadow_ctx;
1408 }
1409
1410 i915_vm_put(ce->vm);
1411 ce->vm = i915_vm_get(&ppgtt->vm);
1412 intel_context_set_single_submission(ce);
1413
1414 /* Max ring buffer size */
1415 if (!intel_uc_wants_guc_submission(&engine->gt->uc))
1416 ce->ring_size = SZ_2M;
1417
1418 s->shadow[i] = ce;
1419 }
1420
1421 bitmap_zero(s->shadow_ctx_desc_updated, I915_NUM_ENGINES);
1422
1423 s->workloads = kmem_cache_create_usercopy("gvt-g_vgpu_workload",
1424 sizeof(struct intel_vgpu_workload), 0,
1425 SLAB_HWCACHE_ALIGN,
1426 offsetof(struct intel_vgpu_workload, rb_tail),
1427 sizeof_field(struct intel_vgpu_workload, rb_tail),
1428 NULL);
1429
1430 if (!s->workloads) {
1431 ret = -ENOMEM;
1432 goto out_shadow_ctx;
1433 }
1434
1435 atomic_set(&s->running_workload_num, 0);
1436 bitmap_zero(s->tlb_handle_pending, I915_NUM_ENGINES);
1437
1438 memset(s->last_ctx, 0, sizeof(s->last_ctx));
1439
1440 i915_vm_put(&ppgtt->vm);
1441 return 0;
1442
1443 out_shadow_ctx:
1444 i915_context_ppgtt_root_restore(s, ppgtt);
1445 for_each_engine(engine, vgpu->gvt->gt, i) {
1446 if (IS_ERR(s->shadow[i]))
1447 break;
1448
1449 intel_context_put(s->shadow[i]);
1450 }
1451 i915_vm_put(&ppgtt->vm);
1452 return ret;
1453 }
1454
1455 /**
1456 * intel_vgpu_select_submission_ops - select virtual submission interface
1457 * @vgpu: a vGPU
1458 * @engine_mask: either ALL_ENGINES or target engine mask
1459 * @interface: expected vGPU virtual submission interface
1460 *
1461 * This function is called when guest configures submission interface.
1462 *
1463 * Returns:
1464 * Zero on success, negative error code if failed.
1465 *
1466 */
intel_vgpu_select_submission_ops(struct intel_vgpu * vgpu,intel_engine_mask_t engine_mask,unsigned int interface)1467 int intel_vgpu_select_submission_ops(struct intel_vgpu *vgpu,
1468 intel_engine_mask_t engine_mask,
1469 unsigned int interface)
1470 {
1471 struct drm_i915_private *i915 = vgpu->gvt->gt->i915;
1472 struct intel_vgpu_submission *s = &vgpu->submission;
1473 const struct intel_vgpu_submission_ops *ops[] = {
1474 [INTEL_VGPU_EXECLIST_SUBMISSION] =
1475 &intel_vgpu_execlist_submission_ops,
1476 };
1477 int ret;
1478
1479 if (drm_WARN_ON(&i915->drm, interface >= ARRAY_SIZE(ops)))
1480 return -EINVAL;
1481
1482 if (drm_WARN_ON(&i915->drm,
1483 interface == 0 && engine_mask != ALL_ENGINES))
1484 return -EINVAL;
1485
1486 if (s->active)
1487 s->ops->clean(vgpu, engine_mask);
1488
1489 if (interface == 0) {
1490 s->ops = NULL;
1491 s->virtual_submission_interface = 0;
1492 s->active = false;
1493 gvt_dbg_core("vgpu%d: remove submission ops\n", vgpu->id);
1494 return 0;
1495 }
1496
1497 ret = ops[interface]->init(vgpu, engine_mask);
1498 if (ret)
1499 return ret;
1500
1501 s->ops = ops[interface];
1502 s->virtual_submission_interface = interface;
1503 s->active = true;
1504
1505 gvt_dbg_core("vgpu%d: activate ops [ %s ]\n",
1506 vgpu->id, s->ops->name);
1507
1508 return 0;
1509 }
1510
1511 /**
1512 * intel_vgpu_destroy_workload - destroy a vGPU workload
1513 * @workload: workload to destroy
1514 *
1515 * This function is called when destroy a vGPU workload.
1516 *
1517 */
intel_vgpu_destroy_workload(struct intel_vgpu_workload * workload)1518 void intel_vgpu_destroy_workload(struct intel_vgpu_workload *workload)
1519 {
1520 struct intel_vgpu_submission *s = &workload->vgpu->submission;
1521
1522 intel_context_unpin(s->shadow[workload->engine->id]);
1523 release_shadow_batch_buffer(workload);
1524 release_shadow_wa_ctx(&workload->wa_ctx);
1525
1526 if (!list_empty(&workload->lri_shadow_mm)) {
1527 struct intel_vgpu_mm *m, *mm;
1528 list_for_each_entry_safe(m, mm, &workload->lri_shadow_mm,
1529 ppgtt_mm.link) {
1530 list_del(&m->ppgtt_mm.link);
1531 intel_vgpu_mm_put(m);
1532 }
1533 }
1534
1535 GEM_BUG_ON(!list_empty(&workload->lri_shadow_mm));
1536 if (workload->shadow_mm)
1537 intel_vgpu_mm_put(workload->shadow_mm);
1538
1539 kmem_cache_free(s->workloads, workload);
1540 }
1541
1542 static struct intel_vgpu_workload *
alloc_workload(struct intel_vgpu * vgpu)1543 alloc_workload(struct intel_vgpu *vgpu)
1544 {
1545 struct intel_vgpu_submission *s = &vgpu->submission;
1546 struct intel_vgpu_workload *workload;
1547
1548 workload = kmem_cache_zalloc(s->workloads, GFP_KERNEL);
1549 if (!workload)
1550 return ERR_PTR(-ENOMEM);
1551
1552 INIT_LIST_HEAD(&workload->list);
1553 INIT_LIST_HEAD(&workload->shadow_bb);
1554 INIT_LIST_HEAD(&workload->lri_shadow_mm);
1555
1556 init_waitqueue_head(&workload->shadow_ctx_status_wq);
1557 atomic_set(&workload->shadow_ctx_active, 0);
1558
1559 workload->status = -EINPROGRESS;
1560 workload->vgpu = vgpu;
1561
1562 return workload;
1563 }
1564
1565 #define RING_CTX_OFF(x) \
1566 offsetof(struct execlist_ring_context, x)
1567
read_guest_pdps(struct intel_vgpu * vgpu,u64 ring_context_gpa,u32 pdp[8])1568 static void read_guest_pdps(struct intel_vgpu *vgpu,
1569 u64 ring_context_gpa, u32 pdp[8])
1570 {
1571 u64 gpa;
1572 int i;
1573
1574 gpa = ring_context_gpa + RING_CTX_OFF(pdps[0].val);
1575
1576 for (i = 0; i < 8; i++)
1577 intel_gvt_read_gpa(vgpu,
1578 gpa + i * 8, &pdp[7 - i], 4);
1579 }
1580
prepare_mm(struct intel_vgpu_workload * workload)1581 static int prepare_mm(struct intel_vgpu_workload *workload)
1582 {
1583 struct execlist_ctx_descriptor_format *desc = &workload->ctx_desc;
1584 struct intel_vgpu_mm *mm;
1585 struct intel_vgpu *vgpu = workload->vgpu;
1586 enum intel_gvt_gtt_type root_entry_type;
1587 u64 pdps[GVT_RING_CTX_NR_PDPS];
1588
1589 switch (desc->addressing_mode) {
1590 case 1: /* legacy 32-bit */
1591 root_entry_type = GTT_TYPE_PPGTT_ROOT_L3_ENTRY;
1592 break;
1593 case 3: /* legacy 64-bit */
1594 root_entry_type = GTT_TYPE_PPGTT_ROOT_L4_ENTRY;
1595 break;
1596 default:
1597 gvt_vgpu_err("Advanced Context mode(SVM) is not supported!\n");
1598 return -EINVAL;
1599 }
1600
1601 read_guest_pdps(workload->vgpu, workload->ring_context_gpa, (void *)pdps);
1602
1603 mm = intel_vgpu_get_ppgtt_mm(workload->vgpu, root_entry_type, pdps);
1604 if (IS_ERR(mm))
1605 return PTR_ERR(mm);
1606
1607 workload->shadow_mm = mm;
1608 return 0;
1609 }
1610
1611 #define same_context(a, b) (((a)->context_id == (b)->context_id) && \
1612 ((a)->lrca == (b)->lrca))
1613
1614 /**
1615 * intel_vgpu_create_workload - create a vGPU workload
1616 * @vgpu: a vGPU
1617 * @engine: the engine
1618 * @desc: a guest context descriptor
1619 *
1620 * This function is called when creating a vGPU workload.
1621 *
1622 * Returns:
1623 * struct intel_vgpu_workload * on success, negative error code in
1624 * pointer if failed.
1625 *
1626 */
1627 struct intel_vgpu_workload *
intel_vgpu_create_workload(struct intel_vgpu * vgpu,const struct intel_engine_cs * engine,struct execlist_ctx_descriptor_format * desc)1628 intel_vgpu_create_workload(struct intel_vgpu *vgpu,
1629 const struct intel_engine_cs *engine,
1630 struct execlist_ctx_descriptor_format *desc)
1631 {
1632 struct intel_vgpu_submission *s = &vgpu->submission;
1633 struct list_head *q = workload_q_head(vgpu, engine);
1634 struct intel_vgpu_workload *last_workload = NULL;
1635 struct intel_vgpu_workload *workload = NULL;
1636 u64 ring_context_gpa;
1637 u32 head, tail, start, ctl, ctx_ctl, per_ctx, indirect_ctx;
1638 u32 guest_head;
1639 int ret;
1640
1641 ring_context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
1642 (u32)((desc->lrca + 1) << I915_GTT_PAGE_SHIFT));
1643 if (ring_context_gpa == INTEL_GVT_INVALID_ADDR) {
1644 gvt_vgpu_err("invalid guest context LRCA: %x\n", desc->lrca);
1645 return ERR_PTR(-EINVAL);
1646 }
1647
1648 intel_gvt_read_gpa(vgpu, ring_context_gpa +
1649 RING_CTX_OFF(ring_header.val), &head, 4);
1650
1651 intel_gvt_read_gpa(vgpu, ring_context_gpa +
1652 RING_CTX_OFF(ring_tail.val), &tail, 4);
1653
1654 guest_head = head;
1655
1656 head &= RB_HEAD_OFF_MASK;
1657 tail &= RB_TAIL_OFF_MASK;
1658
1659 list_for_each_entry_reverse(last_workload, q, list) {
1660
1661 if (same_context(&last_workload->ctx_desc, desc)) {
1662 gvt_dbg_el("ring %s cur workload == last\n",
1663 engine->name);
1664 gvt_dbg_el("ctx head %x real head %lx\n", head,
1665 last_workload->rb_tail);
1666 /*
1667 * cannot use guest context head pointer here,
1668 * as it might not be updated at this time
1669 */
1670 head = last_workload->rb_tail;
1671 break;
1672 }
1673 }
1674
1675 gvt_dbg_el("ring %s begin a new workload\n", engine->name);
1676
1677 /* record some ring buffer register values for scan and shadow */
1678 intel_gvt_read_gpa(vgpu, ring_context_gpa +
1679 RING_CTX_OFF(rb_start.val), &start, 4);
1680 intel_gvt_read_gpa(vgpu, ring_context_gpa +
1681 RING_CTX_OFF(rb_ctrl.val), &ctl, 4);
1682 intel_gvt_read_gpa(vgpu, ring_context_gpa +
1683 RING_CTX_OFF(ctx_ctrl.val), &ctx_ctl, 4);
1684
1685 if (!intel_gvt_ggtt_validate_range(vgpu, start,
1686 _RING_CTL_BUF_SIZE(ctl))) {
1687 gvt_vgpu_err("context contain invalid rb at: 0x%x\n", start);
1688 return ERR_PTR(-EINVAL);
1689 }
1690
1691 workload = alloc_workload(vgpu);
1692 if (IS_ERR(workload))
1693 return workload;
1694
1695 workload->engine = engine;
1696 workload->ctx_desc = *desc;
1697 workload->ring_context_gpa = ring_context_gpa;
1698 workload->rb_head = head;
1699 workload->guest_rb_head = guest_head;
1700 workload->rb_tail = tail;
1701 workload->rb_start = start;
1702 workload->rb_ctl = ctl;
1703
1704 if (engine->id == RCS0) {
1705 intel_gvt_read_gpa(vgpu, ring_context_gpa +
1706 RING_CTX_OFF(bb_per_ctx_ptr.val), &per_ctx, 4);
1707 intel_gvt_read_gpa(vgpu, ring_context_gpa +
1708 RING_CTX_OFF(rcs_indirect_ctx.val), &indirect_ctx, 4);
1709
1710 workload->wa_ctx.indirect_ctx.guest_gma =
1711 indirect_ctx & INDIRECT_CTX_ADDR_MASK;
1712 workload->wa_ctx.indirect_ctx.size =
1713 (indirect_ctx & INDIRECT_CTX_SIZE_MASK) *
1714 CACHELINE_BYTES;
1715
1716 if (workload->wa_ctx.indirect_ctx.size != 0) {
1717 if (!intel_gvt_ggtt_validate_range(vgpu,
1718 workload->wa_ctx.indirect_ctx.guest_gma,
1719 workload->wa_ctx.indirect_ctx.size)) {
1720 gvt_vgpu_err("invalid wa_ctx at: 0x%lx\n",
1721 workload->wa_ctx.indirect_ctx.guest_gma);
1722 kmem_cache_free(s->workloads, workload);
1723 return ERR_PTR(-EINVAL);
1724 }
1725 }
1726
1727 workload->wa_ctx.per_ctx.guest_gma =
1728 per_ctx & PER_CTX_ADDR_MASK;
1729 workload->wa_ctx.per_ctx.valid = per_ctx & 1;
1730 if (workload->wa_ctx.per_ctx.valid) {
1731 if (!intel_gvt_ggtt_validate_range(vgpu,
1732 workload->wa_ctx.per_ctx.guest_gma,
1733 CACHELINE_BYTES)) {
1734 gvt_vgpu_err("invalid per_ctx at: 0x%lx\n",
1735 workload->wa_ctx.per_ctx.guest_gma);
1736 kmem_cache_free(s->workloads, workload);
1737 return ERR_PTR(-EINVAL);
1738 }
1739 }
1740 }
1741
1742 gvt_dbg_el("workload %p ring %s head %x tail %x start %x ctl %x\n",
1743 workload, engine->name, head, tail, start, ctl);
1744
1745 ret = prepare_mm(workload);
1746 if (ret) {
1747 kmem_cache_free(s->workloads, workload);
1748 return ERR_PTR(ret);
1749 }
1750
1751 /* Only scan and shadow the first workload in the queue
1752 * as there is only one pre-allocated buf-obj for shadow.
1753 */
1754 if (list_empty(q)) {
1755 intel_wakeref_t wakeref;
1756
1757 with_intel_runtime_pm(engine->gt->uncore->rpm, wakeref)
1758 ret = intel_gvt_scan_and_shadow_workload(workload);
1759 }
1760
1761 if (ret) {
1762 if (vgpu_is_vm_unhealthy(ret))
1763 enter_failsafe_mode(vgpu, GVT_FAILSAFE_GUEST_ERR);
1764 intel_vgpu_destroy_workload(workload);
1765 return ERR_PTR(ret);
1766 }
1767
1768 ret = intel_context_pin(s->shadow[engine->id]);
1769 if (ret) {
1770 intel_vgpu_destroy_workload(workload);
1771 return ERR_PTR(ret);
1772 }
1773
1774 return workload;
1775 }
1776
1777 /**
1778 * intel_vgpu_queue_workload - Queue a vGPU workload
1779 * @workload: the workload to queue in
1780 */
intel_vgpu_queue_workload(struct intel_vgpu_workload * workload)1781 void intel_vgpu_queue_workload(struct intel_vgpu_workload *workload)
1782 {
1783 list_add_tail(&workload->list,
1784 workload_q_head(workload->vgpu, workload->engine));
1785 intel_gvt_kick_schedule(workload->vgpu->gvt);
1786 wake_up(&workload->vgpu->gvt->scheduler.waitq[workload->engine->id]);
1787 }
1788