• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright © 2017 Intel Corporation
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 shall be included
12  * in all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
15  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
19  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
20  * DEALINGS IN THE SOFTWARE.
21  */
22 
23 /**
24  * @file iris_batch.c
25  *
26  * Batchbuffer and command submission module.
27  *
28  * Every API draw call results in a number of GPU commands, which we
29  * collect into a "batch buffer".  Typically, many draw calls are grouped
30  * into a single batch to amortize command submission overhead.
31  *
32  * We submit batches to the kernel using the I915_GEM_EXECBUFFER2 ioctl.
33  * One critical piece of data is the "validation list", which contains a
34  * list of the buffer objects (BOs) which the commands in the GPU need.
35  * The kernel will make sure these are resident and pinned at the correct
36  * virtual memory address before executing our batch.  If a BO is not in
37  * the validation list, it effectively does not exist, so take care.
38  */
39 
40 #include "iris_batch.h"
41 #include "iris_bufmgr.h"
42 #include "iris_context.h"
43 #include "iris_fence.h"
44 
45 #include "drm-uapi/i915_drm.h"
46 
47 #include "common/gen_aux_map.h"
48 #include "intel/common/gen_gem.h"
49 #include "util/hash_table.h"
50 #include "util/set.h"
51 #include "util/u_upload_mgr.h"
52 #include "main/macros.h"
53 
54 #include <errno.h>
55 #include <xf86drm.h>
56 
57 #if HAVE_VALGRIND
58 #include <valgrind.h>
59 #include <memcheck.h>
60 #define VG(x) x
61 #else
62 #define VG(x)
63 #endif
64 
65 #define FILE_DEBUG_FLAG DEBUG_BUFMGR
66 
67 static void
68 iris_batch_reset(struct iris_batch *batch);
69 
70 static unsigned
num_fences(struct iris_batch * batch)71 num_fences(struct iris_batch *batch)
72 {
73    return util_dynarray_num_elements(&batch->exec_fences,
74                                      struct drm_i915_gem_exec_fence);
75 }
76 
77 /**
78  * Debugging code to dump the fence list, used by INTEL_DEBUG=submit.
79  */
80 static void
dump_fence_list(struct iris_batch * batch)81 dump_fence_list(struct iris_batch *batch)
82 {
83    fprintf(stderr, "Fence list (length %u):      ", num_fences(batch));
84 
85    util_dynarray_foreach(&batch->exec_fences,
86                          struct drm_i915_gem_exec_fence, f) {
87       fprintf(stderr, "%s%u%s ",
88               (f->flags & I915_EXEC_FENCE_WAIT) ? "..." : "",
89               f->handle,
90               (f->flags & I915_EXEC_FENCE_SIGNAL) ? "!" : "");
91    }
92 
93    fprintf(stderr, "\n");
94 }
95 
96 /**
97  * Debugging code to dump the validation list, used by INTEL_DEBUG=submit.
98  */
99 static void
dump_validation_list(struct iris_batch * batch)100 dump_validation_list(struct iris_batch *batch)
101 {
102    fprintf(stderr, "Validation list (length %d):\n", batch->exec_count);
103 
104    for (int i = 0; i < batch->exec_count; i++) {
105       uint64_t flags = batch->validation_list[i].flags;
106       assert(batch->validation_list[i].handle ==
107              batch->exec_bos[i]->gem_handle);
108       fprintf(stderr, "[%2d]: %2d %-14s @ 0x%"PRIx64" (%"PRIu64"B)\t %2d refs %s\n",
109               i,
110               batch->validation_list[i].handle,
111               batch->exec_bos[i]->name,
112               (uint64_t)batch->validation_list[i].offset,
113               batch->exec_bos[i]->size,
114               batch->exec_bos[i]->refcount,
115               (flags & EXEC_OBJECT_WRITE) ? " (write)" : "");
116    }
117 }
118 
119 /**
120  * Return BO information to the batch decoder (for debugging).
121  */
122 static struct gen_batch_decode_bo
decode_get_bo(void * v_batch,bool ppgtt,uint64_t address)123 decode_get_bo(void *v_batch, bool ppgtt, uint64_t address)
124 {
125    struct iris_batch *batch = v_batch;
126 
127    assert(ppgtt);
128 
129    for (int i = 0; i < batch->exec_count; i++) {
130       struct iris_bo *bo = batch->exec_bos[i];
131       /* The decoder zeroes out the top 16 bits, so we need to as well */
132       uint64_t bo_address = bo->gtt_offset & (~0ull >> 16);
133 
134       if (address >= bo_address && address < bo_address + bo->size) {
135          return (struct gen_batch_decode_bo) {
136             .addr = address,
137             .size = bo->size,
138             .map = iris_bo_map(batch->dbg, bo, MAP_READ) +
139                    (address - bo_address),
140          };
141       }
142    }
143 
144    return (struct gen_batch_decode_bo) { };
145 }
146 
147 static unsigned
decode_get_state_size(void * v_batch,uint64_t address,UNUSED uint64_t base_address)148 decode_get_state_size(void *v_batch,
149                       uint64_t address,
150                       UNUSED uint64_t base_address)
151 {
152    struct iris_batch *batch = v_batch;
153    unsigned size = (uintptr_t)
154       _mesa_hash_table_u64_search(batch->state_sizes, address);
155 
156    return size;
157 }
158 
159 /**
160  * Decode the current batch.
161  */
162 static void
decode_batch(struct iris_batch * batch)163 decode_batch(struct iris_batch *batch)
164 {
165    void *map = iris_bo_map(batch->dbg, batch->exec_bos[0], MAP_READ);
166    gen_print_batch(&batch->decoder, map, batch->primary_batch_size,
167                    batch->exec_bos[0]->gtt_offset, false);
168 }
169 
170 void
iris_init_batch(struct iris_context * ice,enum iris_batch_name name,int priority)171 iris_init_batch(struct iris_context *ice,
172                 enum iris_batch_name name,
173                 int priority)
174 {
175    struct iris_batch *batch = &ice->batches[name];
176    struct iris_screen *screen = (void *) ice->ctx.screen;
177 
178    batch->screen = screen;
179    batch->dbg = &ice->dbg;
180    batch->reset = &ice->reset;
181    batch->state_sizes = ice->state.sizes;
182    batch->name = name;
183 
184    batch->fine_fences.uploader =
185       u_upload_create(&ice->ctx, 4096, PIPE_BIND_CUSTOM,
186                       PIPE_USAGE_STAGING, 0);
187    iris_fine_fence_init(batch);
188 
189    batch->hw_ctx_id = iris_create_hw_context(screen->bufmgr);
190    assert(batch->hw_ctx_id);
191 
192    iris_hw_context_set_priority(screen->bufmgr, batch->hw_ctx_id, priority);
193 
194    util_dynarray_init(&batch->exec_fences, ralloc_context(NULL));
195    util_dynarray_init(&batch->syncobjs, ralloc_context(NULL));
196 
197    batch->exec_count = 0;
198    batch->exec_array_size = 100;
199    batch->exec_bos =
200       malloc(batch->exec_array_size * sizeof(batch->exec_bos[0]));
201    batch->validation_list =
202       malloc(batch->exec_array_size * sizeof(batch->validation_list[0]));
203 
204    batch->cache.render = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
205                                                  _mesa_key_pointer_equal);
206 
207    memset(batch->other_batches, 0, sizeof(batch->other_batches));
208 
209    for (int i = 0, j = 0; i < IRIS_BATCH_COUNT; i++) {
210       if (i != name)
211          batch->other_batches[j++] = &ice->batches[i];
212    }
213 
214    if (INTEL_DEBUG) {
215       const unsigned decode_flags =
216          GEN_BATCH_DECODE_FULL |
217          ((INTEL_DEBUG & DEBUG_COLOR) ? GEN_BATCH_DECODE_IN_COLOR : 0) |
218          GEN_BATCH_DECODE_OFFSETS |
219          GEN_BATCH_DECODE_FLOATS;
220 
221       gen_batch_decode_ctx_init(&batch->decoder, &screen->devinfo,
222                                 stderr, decode_flags, NULL,
223                                 decode_get_bo, decode_get_state_size, batch);
224       batch->decoder.dynamic_base = IRIS_MEMZONE_DYNAMIC_START;
225       batch->decoder.instruction_base = IRIS_MEMZONE_SHADER_START;
226       batch->decoder.max_vbo_decoded_lines = 32;
227    }
228 
229    iris_batch_reset(batch);
230 }
231 
232 static struct drm_i915_gem_exec_object2 *
find_validation_entry(struct iris_batch * batch,struct iris_bo * bo)233 find_validation_entry(struct iris_batch *batch, struct iris_bo *bo)
234 {
235    unsigned index = READ_ONCE(bo->index);
236 
237    if (index < batch->exec_count && batch->exec_bos[index] == bo)
238       return &batch->validation_list[index];
239 
240    /* May have been shared between multiple active batches */
241    for (index = 0; index < batch->exec_count; index++) {
242       if (batch->exec_bos[index] == bo)
243          return &batch->validation_list[index];
244    }
245 
246    return NULL;
247 }
248 
249 static void
ensure_exec_obj_space(struct iris_batch * batch,uint32_t count)250 ensure_exec_obj_space(struct iris_batch *batch, uint32_t count)
251 {
252    while (batch->exec_count + count > batch->exec_array_size) {
253       batch->exec_array_size *= 2;
254       batch->exec_bos =
255          realloc(batch->exec_bos,
256                  batch->exec_array_size * sizeof(batch->exec_bos[0]));
257       batch->validation_list =
258          realloc(batch->validation_list,
259                  batch->exec_array_size * sizeof(batch->validation_list[0]));
260    }
261 }
262 
263 /**
264  * Add a buffer to the current batch's validation list.
265  *
266  * You must call this on any BO you wish to use in this batch, to ensure
267  * that it's resident when the GPU commands execute.
268  */
269 void
iris_use_pinned_bo(struct iris_batch * batch,struct iris_bo * bo,bool writable,enum iris_domain access)270 iris_use_pinned_bo(struct iris_batch *batch,
271                    struct iris_bo *bo,
272                    bool writable, enum iris_domain access)
273 {
274    assert(bo->kflags & EXEC_OBJECT_PINNED);
275 
276    /* Never mark the workaround BO with EXEC_OBJECT_WRITE.  We don't care
277     * about the order of any writes to that buffer, and marking it writable
278     * would introduce data dependencies between multiple batches which share
279     * the buffer.
280     */
281    if (bo == batch->screen->workaround_bo)
282       writable = false;
283 
284    if (access < NUM_IRIS_DOMAINS) {
285       assert(batch->sync_region_depth);
286       iris_bo_bump_seqno(bo, batch->next_seqno, access);
287    }
288 
289    struct drm_i915_gem_exec_object2 *existing_entry =
290       find_validation_entry(batch, bo);
291 
292    if (existing_entry) {
293       /* The BO is already in the validation list; mark it writable */
294       if (writable)
295          existing_entry->flags |= EXEC_OBJECT_WRITE;
296 
297       return;
298    }
299 
300    if (bo != batch->bo) {
301       /* This is the first time our batch has seen this BO.  Before we use it,
302        * we may need to flush and synchronize with other batches.
303        */
304       for (int b = 0; b < ARRAY_SIZE(batch->other_batches); b++) {
305          struct drm_i915_gem_exec_object2 *other_entry =
306             find_validation_entry(batch->other_batches[b], bo);
307 
308          /* If the buffer is referenced by another batch, and either batch
309           * intends to write it, then flush the other batch and synchronize.
310           *
311           * Consider these cases:
312           *
313           * 1. They read, we read   =>  No synchronization required.
314           * 2. They read, we write  =>  Synchronize (they need the old value)
315           * 3. They write, we read  =>  Synchronize (we need their new value)
316           * 4. They write, we write =>  Synchronize (order writes)
317           *
318           * The read/read case is very common, as multiple batches usually
319           * share a streaming state buffer or shader assembly buffer, and
320           * we want to avoid synchronizing in this case.
321           */
322          if (other_entry &&
323              ((other_entry->flags & EXEC_OBJECT_WRITE) || writable)) {
324             iris_batch_flush(batch->other_batches[b]);
325             iris_batch_add_syncobj(batch,
326                                    batch->other_batches[b]->last_fence->syncobj,
327                                    I915_EXEC_FENCE_WAIT);
328          }
329       }
330    }
331 
332    /* Now, take a reference and add it to the validation list. */
333    iris_bo_reference(bo);
334 
335    ensure_exec_obj_space(batch, 1);
336 
337    batch->validation_list[batch->exec_count] =
338       (struct drm_i915_gem_exec_object2) {
339          .handle = bo->gem_handle,
340          .offset = bo->gtt_offset,
341          .flags = bo->kflags | (writable ? EXEC_OBJECT_WRITE : 0),
342       };
343 
344    bo->index = batch->exec_count;
345    batch->exec_bos[batch->exec_count] = bo;
346    batch->aperture_space += bo->size;
347 
348    batch->exec_count++;
349 }
350 
351 static void
create_batch(struct iris_batch * batch)352 create_batch(struct iris_batch *batch)
353 {
354    struct iris_screen *screen = batch->screen;
355    struct iris_bufmgr *bufmgr = screen->bufmgr;
356 
357    batch->bo = iris_bo_alloc(bufmgr, "command buffer",
358                              BATCH_SZ + BATCH_RESERVED, IRIS_MEMZONE_OTHER);
359    batch->bo->kflags |= EXEC_OBJECT_CAPTURE;
360    batch->map = iris_bo_map(NULL, batch->bo, MAP_READ | MAP_WRITE);
361    batch->map_next = batch->map;
362 
363    iris_use_pinned_bo(batch, batch->bo, false, IRIS_DOMAIN_NONE);
364 }
365 
366 static void
iris_batch_maybe_noop(struct iris_batch * batch)367 iris_batch_maybe_noop(struct iris_batch *batch)
368 {
369    /* We only insert the NOOP at the beginning of the batch. */
370    assert(iris_batch_bytes_used(batch) == 0);
371 
372    if (batch->noop_enabled) {
373       /* Emit MI_BATCH_BUFFER_END to prevent any further command to be
374        * executed.
375        */
376       uint32_t *map = batch->map_next;
377 
378       map[0] = (0xA << 23);
379 
380       batch->map_next += 4;
381    }
382 }
383 
384 static void
iris_batch_reset(struct iris_batch * batch)385 iris_batch_reset(struct iris_batch *batch)
386 {
387    struct iris_screen *screen = batch->screen;
388 
389    iris_bo_unreference(batch->bo);
390    batch->primary_batch_size = 0;
391    batch->total_chained_batch_size = 0;
392    batch->contains_draw = false;
393    batch->decoder.surface_base = batch->last_surface_base_address;
394 
395    create_batch(batch);
396    assert(batch->bo->index == 0);
397 
398    struct iris_syncobj *syncobj = iris_create_syncobj(screen);
399    iris_batch_add_syncobj(batch, syncobj, I915_EXEC_FENCE_SIGNAL);
400    iris_syncobj_reference(screen, &syncobj, NULL);
401 
402    assert(!batch->sync_region_depth);
403    iris_batch_sync_boundary(batch);
404    iris_batch_mark_reset_sync(batch);
405 
406    /* Always add the workaround BO, it contains a driver identifier at the
407     * beginning quite helpful to debug error states.
408     */
409    iris_use_pinned_bo(batch, screen->workaround_bo, false, IRIS_DOMAIN_NONE);
410 
411    iris_batch_maybe_noop(batch);
412 }
413 
414 void
iris_batch_free(struct iris_batch * batch)415 iris_batch_free(struct iris_batch *batch)
416 {
417    struct iris_screen *screen = batch->screen;
418    struct iris_bufmgr *bufmgr = screen->bufmgr;
419 
420    for (int i = 0; i < batch->exec_count; i++) {
421       iris_bo_unreference(batch->exec_bos[i]);
422    }
423    free(batch->exec_bos);
424    free(batch->validation_list);
425 
426    ralloc_free(batch->exec_fences.mem_ctx);
427 
428    pipe_resource_reference(&batch->fine_fences.ref.res, NULL);
429 
430    util_dynarray_foreach(&batch->syncobjs, struct iris_syncobj *, s)
431       iris_syncobj_reference(screen, s, NULL);
432    ralloc_free(batch->syncobjs.mem_ctx);
433 
434    iris_fine_fence_reference(batch->screen, &batch->last_fence, NULL);
435    u_upload_destroy(batch->fine_fences.uploader);
436 
437    iris_bo_unreference(batch->bo);
438    batch->bo = NULL;
439    batch->map = NULL;
440    batch->map_next = NULL;
441 
442    iris_destroy_hw_context(bufmgr, batch->hw_ctx_id);
443 
444    _mesa_hash_table_destroy(batch->cache.render, NULL);
445 
446    if (INTEL_DEBUG)
447       gen_batch_decode_ctx_finish(&batch->decoder);
448 }
449 
450 /**
451  * If we've chained to a secondary batch, or are getting near to the end,
452  * then flush.  This should only be called between draws.
453  */
454 void
iris_batch_maybe_flush(struct iris_batch * batch,unsigned estimate)455 iris_batch_maybe_flush(struct iris_batch *batch, unsigned estimate)
456 {
457    if (batch->bo != batch->exec_bos[0] ||
458        iris_batch_bytes_used(batch) + estimate >= BATCH_SZ) {
459       iris_batch_flush(batch);
460    }
461 }
462 
463 static void
record_batch_sizes(struct iris_batch * batch)464 record_batch_sizes(struct iris_batch *batch)
465 {
466    unsigned batch_size = iris_batch_bytes_used(batch);
467 
468    VG(VALGRIND_CHECK_MEM_IS_DEFINED(batch->map, batch_size));
469 
470    if (batch->bo == batch->exec_bos[0])
471       batch->primary_batch_size = batch_size;
472 
473    batch->total_chained_batch_size += batch_size;
474 }
475 
476 void
iris_chain_to_new_batch(struct iris_batch * batch)477 iris_chain_to_new_batch(struct iris_batch *batch)
478 {
479    uint32_t *cmd = batch->map_next;
480    uint64_t *addr = batch->map_next + 4;
481    batch->map_next += 12;
482 
483    record_batch_sizes(batch);
484 
485    /* No longer held by batch->bo, still held by validation list */
486    iris_bo_unreference(batch->bo);
487    create_batch(batch);
488 
489    /* Emit MI_BATCH_BUFFER_START to chain to another batch. */
490    *cmd = (0x31 << 23) | (1 << 8) | (3 - 2);
491    *addr = batch->bo->gtt_offset;
492 }
493 
494 static void
add_aux_map_bos_to_batch(struct iris_batch * batch)495 add_aux_map_bos_to_batch(struct iris_batch *batch)
496 {
497    void *aux_map_ctx = iris_bufmgr_get_aux_map_context(batch->screen->bufmgr);
498    if (!aux_map_ctx)
499       return;
500 
501    uint32_t count = gen_aux_map_get_num_buffers(aux_map_ctx);
502    ensure_exec_obj_space(batch, count);
503    gen_aux_map_fill_bos(aux_map_ctx,
504                         (void**)&batch->exec_bos[batch->exec_count], count);
505    for (uint32_t i = 0; i < count; i++) {
506       struct iris_bo *bo = batch->exec_bos[batch->exec_count];
507       iris_bo_reference(bo);
508       batch->validation_list[batch->exec_count] =
509          (struct drm_i915_gem_exec_object2) {
510             .handle = bo->gem_handle,
511             .offset = bo->gtt_offset,
512             .flags = bo->kflags,
513          };
514       batch->aperture_space += bo->size;
515       batch->exec_count++;
516    }
517 }
518 
519 static void
finish_seqno(struct iris_batch * batch)520 finish_seqno(struct iris_batch *batch)
521 {
522    struct iris_fine_fence *sq = iris_fine_fence_new(batch, IRIS_FENCE_END);
523    if (!sq)
524       return;
525 
526    iris_fine_fence_reference(batch->screen, &batch->last_fence, sq);
527    iris_fine_fence_reference(batch->screen, &sq, NULL);
528 }
529 
530 /**
531  * Terminate a batch with MI_BATCH_BUFFER_END.
532  */
533 static void
iris_finish_batch(struct iris_batch * batch)534 iris_finish_batch(struct iris_batch *batch)
535 {
536    add_aux_map_bos_to_batch(batch);
537 
538    finish_seqno(batch);
539 
540    /* Emit MI_BATCH_BUFFER_END to finish our batch. */
541    uint32_t *map = batch->map_next;
542 
543    map[0] = (0xA << 23);
544 
545    batch->map_next += 4;
546 
547    record_batch_sizes(batch);
548 }
549 
550 /**
551  * Replace our current GEM context with a new one (in case it got banned).
552  */
553 static bool
replace_hw_ctx(struct iris_batch * batch)554 replace_hw_ctx(struct iris_batch *batch)
555 {
556    struct iris_screen *screen = batch->screen;
557    struct iris_bufmgr *bufmgr = screen->bufmgr;
558 
559    uint32_t new_ctx = iris_clone_hw_context(bufmgr, batch->hw_ctx_id);
560    if (!new_ctx)
561       return false;
562 
563    iris_destroy_hw_context(bufmgr, batch->hw_ctx_id);
564    batch->hw_ctx_id = new_ctx;
565 
566    /* Notify the context that state must be re-initialized. */
567    iris_lost_context_state(batch);
568 
569    return true;
570 }
571 
572 enum pipe_reset_status
iris_batch_check_for_reset(struct iris_batch * batch)573 iris_batch_check_for_reset(struct iris_batch *batch)
574 {
575    struct iris_screen *screen = batch->screen;
576    enum pipe_reset_status status = PIPE_NO_RESET;
577    struct drm_i915_reset_stats stats = { .ctx_id = batch->hw_ctx_id };
578 
579    if (drmIoctl(screen->fd, DRM_IOCTL_I915_GET_RESET_STATS, &stats))
580       DBG("DRM_IOCTL_I915_GET_RESET_STATS failed: %s\n", strerror(errno));
581 
582    if (stats.batch_active != 0) {
583       /* A reset was observed while a batch from this hardware context was
584        * executing.  Assume that this context was at fault.
585        */
586       status = PIPE_GUILTY_CONTEXT_RESET;
587    } else if (stats.batch_pending != 0) {
588       /* A reset was observed while a batch from this context was in progress,
589        * but the batch was not executing.  In this case, assume that the
590        * context was not at fault.
591        */
592       status = PIPE_INNOCENT_CONTEXT_RESET;
593    }
594 
595    if (status != PIPE_NO_RESET) {
596       /* Our context is likely banned, or at least in an unknown state.
597        * Throw it away and start with a fresh context.  Ideally this may
598        * catch the problem before our next execbuf fails with -EIO.
599        */
600       replace_hw_ctx(batch);
601    }
602 
603    return status;
604 }
605 
606 /**
607  * Submit the batch to the GPU via execbuffer2.
608  */
609 static int
submit_batch(struct iris_batch * batch)610 submit_batch(struct iris_batch *batch)
611 {
612    iris_bo_unmap(batch->bo);
613 
614    /* The requirement for using I915_EXEC_NO_RELOC are:
615     *
616     *   The addresses written in the objects must match the corresponding
617     *   reloc.gtt_offset which in turn must match the corresponding
618     *   execobject.offset.
619     *
620     *   Any render targets written to in the batch must be flagged with
621     *   EXEC_OBJECT_WRITE.
622     *
623     *   To avoid stalling, execobject.offset should match the current
624     *   address of that object within the active context.
625     */
626    struct drm_i915_gem_execbuffer2 execbuf = {
627       .buffers_ptr = (uintptr_t) batch->validation_list,
628       .buffer_count = batch->exec_count,
629       .batch_start_offset = 0,
630       /* This must be QWord aligned. */
631       .batch_len = ALIGN(batch->primary_batch_size, 8),
632       .flags = I915_EXEC_RENDER |
633                I915_EXEC_NO_RELOC |
634                I915_EXEC_BATCH_FIRST |
635                I915_EXEC_HANDLE_LUT,
636       .rsvd1 = batch->hw_ctx_id, /* rsvd1 is actually the context ID */
637    };
638 
639    if (num_fences(batch)) {
640       execbuf.flags |= I915_EXEC_FENCE_ARRAY;
641       execbuf.num_cliprects = num_fences(batch);
642       execbuf.cliprects_ptr =
643          (uintptr_t)util_dynarray_begin(&batch->exec_fences);
644    }
645 
646    int ret = 0;
647    if (!batch->screen->no_hw &&
648        gen_ioctl(batch->screen->fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf))
649       ret = -errno;
650 
651    for (int i = 0; i < batch->exec_count; i++) {
652       struct iris_bo *bo = batch->exec_bos[i];
653 
654       bo->idle = false;
655       bo->index = -1;
656 
657       iris_bo_unreference(bo);
658    }
659 
660    return ret;
661 }
662 
663 static const char *
batch_name_to_string(enum iris_batch_name name)664 batch_name_to_string(enum iris_batch_name name)
665 {
666    const char *names[IRIS_BATCH_COUNT] = {
667       [IRIS_BATCH_RENDER]  = "render",
668       [IRIS_BATCH_COMPUTE] = "compute",
669    };
670    return names[name];
671 }
672 
673 /**
674  * Flush the batch buffer, submitting it to the GPU and resetting it so
675  * we're ready to emit the next batch.
676  */
677 void
_iris_batch_flush(struct iris_batch * batch,const char * file,int line)678 _iris_batch_flush(struct iris_batch *batch, const char *file, int line)
679 {
680    struct iris_screen *screen = batch->screen;
681 
682    if (iris_batch_bytes_used(batch) == 0)
683       return;
684 
685    iris_finish_batch(batch);
686 
687    if (INTEL_DEBUG & (DEBUG_BATCH | DEBUG_SUBMIT | DEBUG_PIPE_CONTROL)) {
688       const char *basefile = strstr(file, "iris/");
689       if (basefile)
690          file = basefile + 5;
691 
692       fprintf(stderr, "%19s:%-3d: %s batch [%u] flush with %5db (%0.1f%%) "
693               "(cmds), %4d BOs (%0.1fMb aperture)\n",
694               file, line, batch_name_to_string(batch->name), batch->hw_ctx_id,
695               batch->total_chained_batch_size,
696               100.0f * batch->total_chained_batch_size / BATCH_SZ,
697               batch->exec_count,
698               (float) batch->aperture_space / (1024 * 1024));
699 
700       if (INTEL_DEBUG & (DEBUG_BATCH | DEBUG_SUBMIT)) {
701          dump_fence_list(batch);
702          dump_validation_list(batch);
703       }
704 
705       if (INTEL_DEBUG & DEBUG_BATCH) {
706          decode_batch(batch);
707       }
708    }
709 
710    int ret = submit_batch(batch);
711 
712    batch->exec_count = 0;
713    batch->aperture_space = 0;
714 
715    util_dynarray_foreach(&batch->syncobjs, struct iris_syncobj *, s)
716       iris_syncobj_reference(screen, s, NULL);
717    util_dynarray_clear(&batch->syncobjs);
718 
719    util_dynarray_clear(&batch->exec_fences);
720 
721    if (INTEL_DEBUG & DEBUG_SYNC) {
722       dbg_printf("waiting for idle\n");
723       iris_bo_wait_rendering(batch->bo); /* if execbuf failed; this is a nop */
724    }
725 
726    /* Start a new batch buffer. */
727    iris_batch_reset(batch);
728 
729    /* EIO means our context is banned.  In this case, try and replace it
730     * with a new logical context, and inform iris_context that all state
731     * has been lost and needs to be re-initialized.  If this succeeds,
732     * dubiously claim success...
733     */
734    if (ret == -EIO && replace_hw_ctx(batch)) {
735       if (batch->reset->reset) {
736          /* Tell gallium frontends the device is lost and it was our fault. */
737          batch->reset->reset(batch->reset->data, PIPE_GUILTY_CONTEXT_RESET);
738       }
739 
740       ret = 0;
741    }
742 
743    if (ret < 0) {
744 #ifdef DEBUG
745       const bool color = INTEL_DEBUG & DEBUG_COLOR;
746       fprintf(stderr, "%siris: Failed to submit batchbuffer: %-80s%s\n",
747               color ? "\e[1;41m" : "", strerror(-ret), color ? "\e[0m" : "");
748 #endif
749       abort();
750    }
751 }
752 
753 /**
754  * Does the current batch refer to the given BO?
755  *
756  * (In other words, is the BO in the current batch's validation list?)
757  */
758 bool
iris_batch_references(struct iris_batch * batch,struct iris_bo * bo)759 iris_batch_references(struct iris_batch *batch, struct iris_bo *bo)
760 {
761    return find_validation_entry(batch, bo) != NULL;
762 }
763 
764 /**
765  * Updates the state of the noop feature.  Returns true if there was a noop
766  * transition that led to state invalidation.
767  */
768 bool
iris_batch_prepare_noop(struct iris_batch * batch,bool noop_enable)769 iris_batch_prepare_noop(struct iris_batch *batch, bool noop_enable)
770 {
771    if (batch->noop_enabled == noop_enable)
772       return 0;
773 
774    batch->noop_enabled = noop_enable;
775 
776    iris_batch_flush(batch);
777 
778    /* If the batch was empty, flush had no effect, so insert our noop. */
779    if (iris_batch_bytes_used(batch) == 0)
780       iris_batch_maybe_noop(batch);
781 
782    /* We only need to update the entire state if we transition from noop ->
783     * not-noop.
784     */
785    return !batch->noop_enabled;
786 }
787