aboutsummaryrefslogtreecommitdiffstats
path: root/drivers/gpu/drm/i915/i915_gem_execbuffer.c
diff options
context:
space:
mode:
Diffstat (limited to 'drivers/gpu/drm/i915/i915_gem_execbuffer.c')
-rw-r--r--drivers/gpu/drm/i915/i915_gem_execbuffer.c401
1 files changed, 243 insertions, 158 deletions
diff --git a/drivers/gpu/drm/i915/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
index bf345777ae9f..885d595e0e02 100644
--- a/drivers/gpu/drm/i915/i915_gem_execbuffer.c
+++ b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
@@ -33,35 +33,35 @@
33#include "intel_drv.h" 33#include "intel_drv.h"
34#include <linux/dma_remapping.h> 34#include <linux/dma_remapping.h>
35 35
36struct eb_objects { 36struct eb_vmas {
37 struct list_head objects; 37 struct list_head vmas;
38 int and; 38 int and;
39 union { 39 union {
40 struct drm_i915_gem_object *lut[0]; 40 struct i915_vma *lut[0];
41 struct hlist_head buckets[0]; 41 struct hlist_head buckets[0];
42 }; 42 };
43}; 43};
44 44
45static struct eb_objects * 45static struct eb_vmas *
46eb_create(struct drm_i915_gem_execbuffer2 *args) 46eb_create(struct drm_i915_gem_execbuffer2 *args, struct i915_address_space *vm)
47{ 47{
48 struct eb_objects *eb = NULL; 48 struct eb_vmas *eb = NULL;
49 49
50 if (args->flags & I915_EXEC_HANDLE_LUT) { 50 if (args->flags & I915_EXEC_HANDLE_LUT) {
51 int size = args->buffer_count; 51 unsigned size = args->buffer_count;
52 size *= sizeof(struct drm_i915_gem_object *); 52 size *= sizeof(struct i915_vma *);
53 size += sizeof(struct eb_objects); 53 size += sizeof(struct eb_vmas);
54 eb = kmalloc(size, GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY); 54 eb = kmalloc(size, GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
55 } 55 }
56 56
57 if (eb == NULL) { 57 if (eb == NULL) {
58 int size = args->buffer_count; 58 unsigned size = args->buffer_count;
59 int count = PAGE_SIZE / sizeof(struct hlist_head) / 2; 59 unsigned count = PAGE_SIZE / sizeof(struct hlist_head) / 2;
60 BUILD_BUG_ON_NOT_POWER_OF_2(PAGE_SIZE / sizeof(struct hlist_head)); 60 BUILD_BUG_ON_NOT_POWER_OF_2(PAGE_SIZE / sizeof(struct hlist_head));
61 while (count > 2*size) 61 while (count > 2*size)
62 count >>= 1; 62 count >>= 1;
63 eb = kzalloc(count*sizeof(struct hlist_head) + 63 eb = kzalloc(count*sizeof(struct hlist_head) +
64 sizeof(struct eb_objects), 64 sizeof(struct eb_vmas),
65 GFP_TEMPORARY); 65 GFP_TEMPORARY);
66 if (eb == NULL) 66 if (eb == NULL)
67 return eb; 67 return eb;
@@ -70,64 +70,102 @@ eb_create(struct drm_i915_gem_execbuffer2 *args)
70 } else 70 } else
71 eb->and = -args->buffer_count; 71 eb->and = -args->buffer_count;
72 72
73 INIT_LIST_HEAD(&eb->objects); 73 INIT_LIST_HEAD(&eb->vmas);
74 return eb; 74 return eb;
75} 75}
76 76
77static void 77static void
78eb_reset(struct eb_objects *eb) 78eb_reset(struct eb_vmas *eb)
79{ 79{
80 if (eb->and >= 0) 80 if (eb->and >= 0)
81 memset(eb->buckets, 0, (eb->and+1)*sizeof(struct hlist_head)); 81 memset(eb->buckets, 0, (eb->and+1)*sizeof(struct hlist_head));
82} 82}
83 83
84static int 84static int
85eb_lookup_objects(struct eb_objects *eb, 85eb_lookup_vmas(struct eb_vmas *eb,
86 struct drm_i915_gem_exec_object2 *exec, 86 struct drm_i915_gem_exec_object2 *exec,
87 const struct drm_i915_gem_execbuffer2 *args, 87 const struct drm_i915_gem_execbuffer2 *args,
88 struct drm_file *file) 88 struct i915_address_space *vm,
89 struct drm_file *file)
89{ 90{
90 int i; 91 struct drm_i915_gem_object *obj;
92 struct list_head objects;
93 int i, ret = 0;
91 94
95 INIT_LIST_HEAD(&objects);
92 spin_lock(&file->table_lock); 96 spin_lock(&file->table_lock);
97 /* Grab a reference to the object and release the lock so we can lookup
98 * or create the VMA without using GFP_ATOMIC */
93 for (i = 0; i < args->buffer_count; i++) { 99 for (i = 0; i < args->buffer_count; i++) {
94 struct drm_i915_gem_object *obj;
95
96 obj = to_intel_bo(idr_find(&file->object_idr, exec[i].handle)); 100 obj = to_intel_bo(idr_find(&file->object_idr, exec[i].handle));
97 if (obj == NULL) { 101 if (obj == NULL) {
98 spin_unlock(&file->table_lock); 102 spin_unlock(&file->table_lock);
99 DRM_DEBUG("Invalid object handle %d at index %d\n", 103 DRM_DEBUG("Invalid object handle %d at index %d\n",
100 exec[i].handle, i); 104 exec[i].handle, i);
101 return -ENOENT; 105 ret = -ENOENT;
106 goto out;
102 } 107 }
103 108
104 if (!list_empty(&obj->exec_list)) { 109 if (!list_empty(&obj->obj_exec_link)) {
105 spin_unlock(&file->table_lock); 110 spin_unlock(&file->table_lock);
106 DRM_DEBUG("Object %p [handle %d, index %d] appears more than once in object list\n", 111 DRM_DEBUG("Object %p [handle %d, index %d] appears more than once in object list\n",
107 obj, exec[i].handle, i); 112 obj, exec[i].handle, i);
108 return -EINVAL; 113 ret = -EINVAL;
114 goto out;
109 } 115 }
110 116
111 drm_gem_object_reference(&obj->base); 117 drm_gem_object_reference(&obj->base);
112 list_add_tail(&obj->exec_list, &eb->objects); 118 list_add_tail(&obj->obj_exec_link, &objects);
119 }
120 spin_unlock(&file->table_lock);
113 121
114 obj->exec_entry = &exec[i]; 122 i = 0;
123 list_for_each_entry(obj, &objects, obj_exec_link) {
124 struct i915_vma *vma;
125
126 /*
127 * NOTE: We can leak any vmas created here when something fails
128 * later on. But that's no issue since vma_unbind can deal with
129 * vmas which are not actually bound. And since only
130 * lookup_or_create exists as an interface to get at the vma
131 * from the (obj, vm) we don't run the risk of creating
132 * duplicated vmas for the same vm.
133 */
134 vma = i915_gem_obj_lookup_or_create_vma(obj, vm);
135 if (IS_ERR(vma)) {
136 DRM_DEBUG("Failed to lookup VMA\n");
137 ret = PTR_ERR(vma);
138 goto out;
139 }
140
141 list_add_tail(&vma->exec_list, &eb->vmas);
142
143 vma->exec_entry = &exec[i];
115 if (eb->and < 0) { 144 if (eb->and < 0) {
116 eb->lut[i] = obj; 145 eb->lut[i] = vma;
117 } else { 146 } else {
118 uint32_t handle = args->flags & I915_EXEC_HANDLE_LUT ? i : exec[i].handle; 147 uint32_t handle = args->flags & I915_EXEC_HANDLE_LUT ? i : exec[i].handle;
119 obj->exec_handle = handle; 148 vma->exec_handle = handle;
120 hlist_add_head(&obj->exec_node, 149 hlist_add_head(&vma->exec_node,
121 &eb->buckets[handle & eb->and]); 150 &eb->buckets[handle & eb->and]);
122 } 151 }
152 ++i;
123 } 153 }
124 spin_unlock(&file->table_lock);
125 154
126 return 0; 155
156out:
157 while (!list_empty(&objects)) {
158 obj = list_first_entry(&objects,
159 struct drm_i915_gem_object,
160 obj_exec_link);
161 list_del_init(&obj->obj_exec_link);
162 if (ret)
163 drm_gem_object_unreference(&obj->base);
164 }
165 return ret;
127} 166}
128 167
129static struct drm_i915_gem_object * 168static struct i915_vma *eb_get_vma(struct eb_vmas *eb, unsigned long handle)
130eb_get_object(struct eb_objects *eb, unsigned long handle)
131{ 169{
132 if (eb->and < 0) { 170 if (eb->and < 0) {
133 if (handle >= -eb->and) 171 if (handle >= -eb->and)
@@ -139,34 +177,33 @@ eb_get_object(struct eb_objects *eb, unsigned long handle)
139 177
140 head = &eb->buckets[handle & eb->and]; 178 head = &eb->buckets[handle & eb->and];
141 hlist_for_each(node, head) { 179 hlist_for_each(node, head) {
142 struct drm_i915_gem_object *obj; 180 struct i915_vma *vma;
143 181
144 obj = hlist_entry(node, struct drm_i915_gem_object, exec_node); 182 vma = hlist_entry(node, struct i915_vma, exec_node);
145 if (obj->exec_handle == handle) 183 if (vma->exec_handle == handle)
146 return obj; 184 return vma;
147 } 185 }
148 return NULL; 186 return NULL;
149 } 187 }
150} 188}
151 189
152static void 190static void eb_destroy(struct eb_vmas *eb) {
153eb_destroy(struct eb_objects *eb) 191 while (!list_empty(&eb->vmas)) {
154{ 192 struct i915_vma *vma;
155 while (!list_empty(&eb->objects)) {
156 struct drm_i915_gem_object *obj;
157 193
158 obj = list_first_entry(&eb->objects, 194 vma = list_first_entry(&eb->vmas,
159 struct drm_i915_gem_object, 195 struct i915_vma,
160 exec_list); 196 exec_list);
161 list_del_init(&obj->exec_list); 197 list_del_init(&vma->exec_list);
162 drm_gem_object_unreference(&obj->base); 198 drm_gem_object_unreference(&vma->obj->base);
163 } 199 }
164 kfree(eb); 200 kfree(eb);
165} 201}
166 202
167static inline int use_cpu_reloc(struct drm_i915_gem_object *obj) 203static inline int use_cpu_reloc(struct drm_i915_gem_object *obj)
168{ 204{
169 return (obj->base.write_domain == I915_GEM_DOMAIN_CPU || 205 return (HAS_LLC(obj->base.dev) ||
206 obj->base.write_domain == I915_GEM_DOMAIN_CPU ||
170 !obj->map_and_fenceable || 207 !obj->map_and_fenceable ||
171 obj->cache_level != I915_CACHE_NONE); 208 obj->cache_level != I915_CACHE_NONE);
172} 209}
@@ -175,17 +212,31 @@ static int
175relocate_entry_cpu(struct drm_i915_gem_object *obj, 212relocate_entry_cpu(struct drm_i915_gem_object *obj,
176 struct drm_i915_gem_relocation_entry *reloc) 213 struct drm_i915_gem_relocation_entry *reloc)
177{ 214{
215 struct drm_device *dev = obj->base.dev;
178 uint32_t page_offset = offset_in_page(reloc->offset); 216 uint32_t page_offset = offset_in_page(reloc->offset);
179 char *vaddr; 217 char *vaddr;
180 int ret = -EINVAL; 218 int ret = -EINVAL;
181 219
182 ret = i915_gem_object_set_to_cpu_domain(obj, 1); 220 ret = i915_gem_object_set_to_cpu_domain(obj, true);
183 if (ret) 221 if (ret)
184 return ret; 222 return ret;
185 223
186 vaddr = kmap_atomic(i915_gem_object_get_page(obj, 224 vaddr = kmap_atomic(i915_gem_object_get_page(obj,
187 reloc->offset >> PAGE_SHIFT)); 225 reloc->offset >> PAGE_SHIFT));
188 *(uint32_t *)(vaddr + page_offset) = reloc->delta; 226 *(uint32_t *)(vaddr + page_offset) = reloc->delta;
227
228 if (INTEL_INFO(dev)->gen >= 8) {
229 page_offset = offset_in_page(page_offset + sizeof(uint32_t));
230
231 if (page_offset == 0) {
232 kunmap_atomic(vaddr);
233 vaddr = kmap_atomic(i915_gem_object_get_page(obj,
234 (reloc->offset + sizeof(uint32_t)) >> PAGE_SHIFT));
235 }
236
237 *(uint32_t *)(vaddr + page_offset) = 0;
238 }
239
189 kunmap_atomic(vaddr); 240 kunmap_atomic(vaddr);
190 241
191 return 0; 242 return 0;
@@ -216,6 +267,21 @@ relocate_entry_gtt(struct drm_i915_gem_object *obj,
216 reloc_entry = (uint32_t __iomem *) 267 reloc_entry = (uint32_t __iomem *)
217 (reloc_page + offset_in_page(reloc->offset)); 268 (reloc_page + offset_in_page(reloc->offset));
218 iowrite32(reloc->delta, reloc_entry); 269 iowrite32(reloc->delta, reloc_entry);
270
271 if (INTEL_INFO(dev)->gen >= 8) {
272 reloc_entry += 1;
273
274 if (offset_in_page(reloc->offset + sizeof(uint32_t)) == 0) {
275 io_mapping_unmap_atomic(reloc_page);
276 reloc_page = io_mapping_map_atomic_wc(
277 dev_priv->gtt.mappable,
278 reloc->offset + sizeof(uint32_t));
279 reloc_entry = reloc_page;
280 }
281
282 iowrite32(0, reloc_entry);
283 }
284
219 io_mapping_unmap_atomic(reloc_page); 285 io_mapping_unmap_atomic(reloc_page);
220 286
221 return 0; 287 return 0;
@@ -223,22 +289,24 @@ relocate_entry_gtt(struct drm_i915_gem_object *obj,
223 289
224static int 290static int
225i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj, 291i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
226 struct eb_objects *eb, 292 struct eb_vmas *eb,
227 struct drm_i915_gem_relocation_entry *reloc, 293 struct drm_i915_gem_relocation_entry *reloc,
228 struct i915_address_space *vm) 294 struct i915_address_space *vm)
229{ 295{
230 struct drm_device *dev = obj->base.dev; 296 struct drm_device *dev = obj->base.dev;
231 struct drm_gem_object *target_obj; 297 struct drm_gem_object *target_obj;
232 struct drm_i915_gem_object *target_i915_obj; 298 struct drm_i915_gem_object *target_i915_obj;
299 struct i915_vma *target_vma;
233 uint32_t target_offset; 300 uint32_t target_offset;
234 int ret = -EINVAL; 301 int ret = -EINVAL;
235 302
236 /* we've already hold a reference to all valid objects */ 303 /* we've already hold a reference to all valid objects */
237 target_obj = &eb_get_object(eb, reloc->target_handle)->base; 304 target_vma = eb_get_vma(eb, reloc->target_handle);
238 if (unlikely(target_obj == NULL)) 305 if (unlikely(target_vma == NULL))
239 return -ENOENT; 306 return -ENOENT;
307 target_i915_obj = target_vma->obj;
308 target_obj = &target_vma->obj->base;
240 309
241 target_i915_obj = to_intel_bo(target_obj);
242 target_offset = i915_gem_obj_ggtt_offset(target_i915_obj); 310 target_offset = i915_gem_obj_ggtt_offset(target_i915_obj);
243 311
244 /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and 312 /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
@@ -284,7 +352,8 @@ i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
284 return 0; 352 return 0;
285 353
286 /* Check that the relocation address is valid... */ 354 /* Check that the relocation address is valid... */
287 if (unlikely(reloc->offset > obj->base.size - 4)) { 355 if (unlikely(reloc->offset >
356 obj->base.size - (INTEL_INFO(dev)->gen >= 8 ? 8 : 4))) {
288 DRM_DEBUG("Relocation beyond object bounds: " 357 DRM_DEBUG("Relocation beyond object bounds: "
289 "obj %p target %d offset %d size %d.\n", 358 "obj %p target %d offset %d size %d.\n",
290 obj, reloc->target_handle, 359 obj, reloc->target_handle,
@@ -320,14 +389,13 @@ i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
320} 389}
321 390
322static int 391static int
323i915_gem_execbuffer_relocate_object(struct drm_i915_gem_object *obj, 392i915_gem_execbuffer_relocate_vma(struct i915_vma *vma,
324 struct eb_objects *eb, 393 struct eb_vmas *eb)
325 struct i915_address_space *vm)
326{ 394{
327#define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry)) 395#define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
328 struct drm_i915_gem_relocation_entry stack_reloc[N_RELOC(512)]; 396 struct drm_i915_gem_relocation_entry stack_reloc[N_RELOC(512)];
329 struct drm_i915_gem_relocation_entry __user *user_relocs; 397 struct drm_i915_gem_relocation_entry __user *user_relocs;
330 struct drm_i915_gem_exec_object2 *entry = obj->exec_entry; 398 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
331 int remain, ret; 399 int remain, ret;
332 400
333 user_relocs = to_user_ptr(entry->relocs_ptr); 401 user_relocs = to_user_ptr(entry->relocs_ptr);
@@ -346,8 +414,8 @@ i915_gem_execbuffer_relocate_object(struct drm_i915_gem_object *obj,
346 do { 414 do {
347 u64 offset = r->presumed_offset; 415 u64 offset = r->presumed_offset;
348 416
349 ret = i915_gem_execbuffer_relocate_entry(obj, eb, r, 417 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, r,
350 vm); 418 vma->vm);
351 if (ret) 419 if (ret)
352 return ret; 420 return ret;
353 421
@@ -368,17 +436,16 @@ i915_gem_execbuffer_relocate_object(struct drm_i915_gem_object *obj,
368} 436}
369 437
370static int 438static int
371i915_gem_execbuffer_relocate_object_slow(struct drm_i915_gem_object *obj, 439i915_gem_execbuffer_relocate_vma_slow(struct i915_vma *vma,
372 struct eb_objects *eb, 440 struct eb_vmas *eb,
373 struct drm_i915_gem_relocation_entry *relocs, 441 struct drm_i915_gem_relocation_entry *relocs)
374 struct i915_address_space *vm)
375{ 442{
376 const struct drm_i915_gem_exec_object2 *entry = obj->exec_entry; 443 const struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
377 int i, ret; 444 int i, ret;
378 445
379 for (i = 0; i < entry->relocation_count; i++) { 446 for (i = 0; i < entry->relocation_count; i++) {
380 ret = i915_gem_execbuffer_relocate_entry(obj, eb, &relocs[i], 447 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, &relocs[i],
381 vm); 448 vma->vm);
382 if (ret) 449 if (ret)
383 return ret; 450 return ret;
384 } 451 }
@@ -387,10 +454,10 @@ i915_gem_execbuffer_relocate_object_slow(struct drm_i915_gem_object *obj,
387} 454}
388 455
389static int 456static int
390i915_gem_execbuffer_relocate(struct eb_objects *eb, 457i915_gem_execbuffer_relocate(struct eb_vmas *eb,
391 struct i915_address_space *vm) 458 struct i915_address_space *vm)
392{ 459{
393 struct drm_i915_gem_object *obj; 460 struct i915_vma *vma;
394 int ret = 0; 461 int ret = 0;
395 462
396 /* This is the fast path and we cannot handle a pagefault whilst 463 /* This is the fast path and we cannot handle a pagefault whilst
@@ -401,8 +468,8 @@ i915_gem_execbuffer_relocate(struct eb_objects *eb,
401 * lockdep complains vehemently. 468 * lockdep complains vehemently.
402 */ 469 */
403 pagefault_disable(); 470 pagefault_disable();
404 list_for_each_entry(obj, &eb->objects, exec_list) { 471 list_for_each_entry(vma, &eb->vmas, exec_list) {
405 ret = i915_gem_execbuffer_relocate_object(obj, eb, vm); 472 ret = i915_gem_execbuffer_relocate_vma(vma, eb);
406 if (ret) 473 if (ret)
407 break; 474 break;
408 } 475 }
@@ -415,31 +482,32 @@ i915_gem_execbuffer_relocate(struct eb_objects *eb,
415#define __EXEC_OBJECT_HAS_FENCE (1<<30) 482#define __EXEC_OBJECT_HAS_FENCE (1<<30)
416 483
417static int 484static int
418need_reloc_mappable(struct drm_i915_gem_object *obj) 485need_reloc_mappable(struct i915_vma *vma)
419{ 486{
420 struct drm_i915_gem_exec_object2 *entry = obj->exec_entry; 487 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
421 return entry->relocation_count && !use_cpu_reloc(obj); 488 return entry->relocation_count && !use_cpu_reloc(vma->obj) &&
489 i915_is_ggtt(vma->vm);
422} 490}
423 491
424static int 492static int
425i915_gem_execbuffer_reserve_object(struct drm_i915_gem_object *obj, 493i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,
426 struct intel_ring_buffer *ring, 494 struct intel_ring_buffer *ring,
427 struct i915_address_space *vm, 495 bool *need_reloc)
428 bool *need_reloc)
429{ 496{
430 struct drm_i915_private *dev_priv = obj->base.dev->dev_private; 497 struct drm_i915_private *dev_priv = ring->dev->dev_private;
431 struct drm_i915_gem_exec_object2 *entry = obj->exec_entry; 498 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
432 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4; 499 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
433 bool need_fence, need_mappable; 500 bool need_fence, need_mappable;
501 struct drm_i915_gem_object *obj = vma->obj;
434 int ret; 502 int ret;
435 503
436 need_fence = 504 need_fence =
437 has_fenced_gpu_access && 505 has_fenced_gpu_access &&
438 entry->flags & EXEC_OBJECT_NEEDS_FENCE && 506 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
439 obj->tiling_mode != I915_TILING_NONE; 507 obj->tiling_mode != I915_TILING_NONE;
440 need_mappable = need_fence || need_reloc_mappable(obj); 508 need_mappable = need_fence || need_reloc_mappable(vma);
441 509
442 ret = i915_gem_object_pin(obj, vm, entry->alignment, need_mappable, 510 ret = i915_gem_object_pin(obj, vma->vm, entry->alignment, need_mappable,
443 false); 511 false);
444 if (ret) 512 if (ret)
445 return ret; 513 return ret;
@@ -467,8 +535,8 @@ i915_gem_execbuffer_reserve_object(struct drm_i915_gem_object *obj,
467 obj->has_aliasing_ppgtt_mapping = 1; 535 obj->has_aliasing_ppgtt_mapping = 1;
468 } 536 }
469 537
470 if (entry->offset != i915_gem_obj_offset(obj, vm)) { 538 if (entry->offset != vma->node.start) {
471 entry->offset = i915_gem_obj_offset(obj, vm); 539 entry->offset = vma->node.start;
472 *need_reloc = true; 540 *need_reloc = true;
473 } 541 }
474 542
@@ -485,14 +553,15 @@ i915_gem_execbuffer_reserve_object(struct drm_i915_gem_object *obj,
485} 553}
486 554
487static void 555static void
488i915_gem_execbuffer_unreserve_object(struct drm_i915_gem_object *obj) 556i915_gem_execbuffer_unreserve_vma(struct i915_vma *vma)
489{ 557{
490 struct drm_i915_gem_exec_object2 *entry; 558 struct drm_i915_gem_exec_object2 *entry;
559 struct drm_i915_gem_object *obj = vma->obj;
491 560
492 if (!i915_gem_obj_bound_any(obj)) 561 if (!drm_mm_node_allocated(&vma->node))
493 return; 562 return;
494 563
495 entry = obj->exec_entry; 564 entry = vma->exec_entry;
496 565
497 if (entry->flags & __EXEC_OBJECT_HAS_FENCE) 566 if (entry->flags & __EXEC_OBJECT_HAS_FENCE)
498 i915_gem_object_unpin_fence(obj); 567 i915_gem_object_unpin_fence(obj);
@@ -505,41 +574,46 @@ i915_gem_execbuffer_unreserve_object(struct drm_i915_gem_object *obj)
505 574
506static int 575static int
507i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring, 576i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
508 struct list_head *objects, 577 struct list_head *vmas,
509 struct i915_address_space *vm,
510 bool *need_relocs) 578 bool *need_relocs)
511{ 579{
512 struct drm_i915_gem_object *obj; 580 struct drm_i915_gem_object *obj;
513 struct list_head ordered_objects; 581 struct i915_vma *vma;
582 struct i915_address_space *vm;
583 struct list_head ordered_vmas;
514 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4; 584 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
515 int retry; 585 int retry;
516 586
517 INIT_LIST_HEAD(&ordered_objects); 587 if (list_empty(vmas))
518 while (!list_empty(objects)) { 588 return 0;
589
590 vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm;
591
592 INIT_LIST_HEAD(&ordered_vmas);
593 while (!list_empty(vmas)) {
519 struct drm_i915_gem_exec_object2 *entry; 594 struct drm_i915_gem_exec_object2 *entry;
520 bool need_fence, need_mappable; 595 bool need_fence, need_mappable;
521 596
522 obj = list_first_entry(objects, 597 vma = list_first_entry(vmas, struct i915_vma, exec_list);
523 struct drm_i915_gem_object, 598 obj = vma->obj;
524 exec_list); 599 entry = vma->exec_entry;
525 entry = obj->exec_entry;
526 600
527 need_fence = 601 need_fence =
528 has_fenced_gpu_access && 602 has_fenced_gpu_access &&
529 entry->flags & EXEC_OBJECT_NEEDS_FENCE && 603 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
530 obj->tiling_mode != I915_TILING_NONE; 604 obj->tiling_mode != I915_TILING_NONE;
531 need_mappable = need_fence || need_reloc_mappable(obj); 605 need_mappable = need_fence || need_reloc_mappable(vma);
532 606
533 if (need_mappable) 607 if (need_mappable)
534 list_move(&obj->exec_list, &ordered_objects); 608 list_move(&vma->exec_list, &ordered_vmas);
535 else 609 else
536 list_move_tail(&obj->exec_list, &ordered_objects); 610 list_move_tail(&vma->exec_list, &ordered_vmas);
537 611
538 obj->base.pending_read_domains = I915_GEM_GPU_DOMAINS & ~I915_GEM_DOMAIN_COMMAND; 612 obj->base.pending_read_domains = I915_GEM_GPU_DOMAINS & ~I915_GEM_DOMAIN_COMMAND;
539 obj->base.pending_write_domain = 0; 613 obj->base.pending_write_domain = 0;
540 obj->pending_fenced_gpu_access = false; 614 obj->pending_fenced_gpu_access = false;
541 } 615 }
542 list_splice(&ordered_objects, objects); 616 list_splice(&ordered_vmas, vmas);
543 617
544 /* Attempt to pin all of the buffers into the GTT. 618 /* Attempt to pin all of the buffers into the GTT.
545 * This is done in 3 phases: 619 * This is done in 3 phases:
@@ -558,52 +632,52 @@ i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
558 int ret = 0; 632 int ret = 0;
559 633
560 /* Unbind any ill-fitting objects or pin. */ 634 /* Unbind any ill-fitting objects or pin. */
561 list_for_each_entry(obj, objects, exec_list) { 635 list_for_each_entry(vma, vmas, exec_list) {
562 struct drm_i915_gem_exec_object2 *entry = obj->exec_entry; 636 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
563 bool need_fence, need_mappable; 637 bool need_fence, need_mappable;
564 u32 obj_offset;
565 638
566 if (!i915_gem_obj_bound(obj, vm)) 639 obj = vma->obj;
640
641 if (!drm_mm_node_allocated(&vma->node))
567 continue; 642 continue;
568 643
569 obj_offset = i915_gem_obj_offset(obj, vm);
570 need_fence = 644 need_fence =
571 has_fenced_gpu_access && 645 has_fenced_gpu_access &&
572 entry->flags & EXEC_OBJECT_NEEDS_FENCE && 646 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
573 obj->tiling_mode != I915_TILING_NONE; 647 obj->tiling_mode != I915_TILING_NONE;
574 need_mappable = need_fence || need_reloc_mappable(obj); 648 need_mappable = need_fence || need_reloc_mappable(vma);
575 649
576 WARN_ON((need_mappable || need_fence) && 650 WARN_ON((need_mappable || need_fence) &&
577 !i915_is_ggtt(vm)); 651 !i915_is_ggtt(vma->vm));
578 652
579 if ((entry->alignment && 653 if ((entry->alignment &&
580 obj_offset & (entry->alignment - 1)) || 654 vma->node.start & (entry->alignment - 1)) ||
581 (need_mappable && !obj->map_and_fenceable)) 655 (need_mappable && !obj->map_and_fenceable))
582 ret = i915_vma_unbind(i915_gem_obj_to_vma(obj, vm)); 656 ret = i915_vma_unbind(vma);
583 else 657 else
584 ret = i915_gem_execbuffer_reserve_object(obj, ring, vm, need_relocs); 658 ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
585 if (ret) 659 if (ret)
586 goto err; 660 goto err;
587 } 661 }
588 662
589 /* Bind fresh objects */ 663 /* Bind fresh objects */
590 list_for_each_entry(obj, objects, exec_list) { 664 list_for_each_entry(vma, vmas, exec_list) {
591 if (i915_gem_obj_bound(obj, vm)) 665 if (drm_mm_node_allocated(&vma->node))
592 continue; 666 continue;
593 667
594 ret = i915_gem_execbuffer_reserve_object(obj, ring, vm, need_relocs); 668 ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
595 if (ret) 669 if (ret)
596 goto err; 670 goto err;
597 } 671 }
598 672
599err: /* Decrement pin count for bound objects */ 673err: /* Decrement pin count for bound objects */
600 list_for_each_entry(obj, objects, exec_list) 674 list_for_each_entry(vma, vmas, exec_list)
601 i915_gem_execbuffer_unreserve_object(obj); 675 i915_gem_execbuffer_unreserve_vma(vma);
602 676
603 if (ret != -ENOSPC || retry++) 677 if (ret != -ENOSPC || retry++)
604 return ret; 678 return ret;
605 679
606 ret = i915_gem_evict_everything(ring->dev); 680 ret = i915_gem_evict_vm(vm, true);
607 if (ret) 681 if (ret)
608 return ret; 682 return ret;
609 } while (1); 683 } while (1);
@@ -614,24 +688,27 @@ i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
614 struct drm_i915_gem_execbuffer2 *args, 688 struct drm_i915_gem_execbuffer2 *args,
615 struct drm_file *file, 689 struct drm_file *file,
616 struct intel_ring_buffer *ring, 690 struct intel_ring_buffer *ring,
617 struct eb_objects *eb, 691 struct eb_vmas *eb,
618 struct drm_i915_gem_exec_object2 *exec, 692 struct drm_i915_gem_exec_object2 *exec)
619 struct i915_address_space *vm)
620{ 693{
621 struct drm_i915_gem_relocation_entry *reloc; 694 struct drm_i915_gem_relocation_entry *reloc;
622 struct drm_i915_gem_object *obj; 695 struct i915_address_space *vm;
696 struct i915_vma *vma;
623 bool need_relocs; 697 bool need_relocs;
624 int *reloc_offset; 698 int *reloc_offset;
625 int i, total, ret; 699 int i, total, ret;
626 int count = args->buffer_count; 700 unsigned count = args->buffer_count;
701
702 if (WARN_ON(list_empty(&eb->vmas)))
703 return 0;
704
705 vm = list_first_entry(&eb->vmas, struct i915_vma, exec_list)->vm;
627 706
628 /* We may process another execbuffer during the unlock... */ 707 /* We may process another execbuffer during the unlock... */
629 while (!list_empty(&eb->objects)) { 708 while (!list_empty(&eb->vmas)) {
630 obj = list_first_entry(&eb->objects, 709 vma = list_first_entry(&eb->vmas, struct i915_vma, exec_list);
631 struct drm_i915_gem_object, 710 list_del_init(&vma->exec_list);
632 exec_list); 711 drm_gem_object_unreference(&vma->obj->base);
633 list_del_init(&obj->exec_list);
634 drm_gem_object_unreference(&obj->base);
635 } 712 }
636 713
637 mutex_unlock(&dev->struct_mutex); 714 mutex_unlock(&dev->struct_mutex);
@@ -695,20 +772,19 @@ i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
695 772
696 /* reacquire the objects */ 773 /* reacquire the objects */
697 eb_reset(eb); 774 eb_reset(eb);
698 ret = eb_lookup_objects(eb, exec, args, file); 775 ret = eb_lookup_vmas(eb, exec, args, vm, file);
699 if (ret) 776 if (ret)
700 goto err; 777 goto err;
701 778
702 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0; 779 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
703 ret = i915_gem_execbuffer_reserve(ring, &eb->objects, vm, &need_relocs); 780 ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, &need_relocs);
704 if (ret) 781 if (ret)
705 goto err; 782 goto err;
706 783
707 list_for_each_entry(obj, &eb->objects, exec_list) { 784 list_for_each_entry(vma, &eb->vmas, exec_list) {
708 int offset = obj->exec_entry - exec; 785 int offset = vma->exec_entry - exec;
709 ret = i915_gem_execbuffer_relocate_object_slow(obj, eb, 786 ret = i915_gem_execbuffer_relocate_vma_slow(vma, eb,
710 reloc + reloc_offset[offset], 787 reloc + reloc_offset[offset]);
711 vm);
712 if (ret) 788 if (ret)
713 goto err; 789 goto err;
714 } 790 }
@@ -727,14 +803,15 @@ err:
727 803
728static int 804static int
729i915_gem_execbuffer_move_to_gpu(struct intel_ring_buffer *ring, 805i915_gem_execbuffer_move_to_gpu(struct intel_ring_buffer *ring,
730 struct list_head *objects) 806 struct list_head *vmas)
731{ 807{
732 struct drm_i915_gem_object *obj; 808 struct i915_vma *vma;
733 uint32_t flush_domains = 0; 809 uint32_t flush_domains = 0;
734 bool flush_chipset = false; 810 bool flush_chipset = false;
735 int ret; 811 int ret;
736 812
737 list_for_each_entry(obj, objects, exec_list) { 813 list_for_each_entry(vma, vmas, exec_list) {
814 struct drm_i915_gem_object *obj = vma->obj;
738 ret = i915_gem_object_sync(obj, ring); 815 ret = i915_gem_object_sync(obj, ring);
739 if (ret) 816 if (ret)
740 return ret; 817 return ret;
@@ -771,8 +848,8 @@ validate_exec_list(struct drm_i915_gem_exec_object2 *exec,
771 int count) 848 int count)
772{ 849{
773 int i; 850 int i;
774 int relocs_total = 0; 851 unsigned relocs_total = 0;
775 int relocs_max = INT_MAX / sizeof(struct drm_i915_gem_relocation_entry); 852 unsigned relocs_max = UINT_MAX / sizeof(struct drm_i915_gem_relocation_entry);
776 853
777 for (i = 0; i < count; i++) { 854 for (i = 0; i < count; i++) {
778 char __user *ptr = to_user_ptr(exec[i].relocs_ptr); 855 char __user *ptr = to_user_ptr(exec[i].relocs_ptr);
@@ -809,13 +886,13 @@ validate_exec_list(struct drm_i915_gem_exec_object2 *exec,
809} 886}
810 887
811static void 888static void
812i915_gem_execbuffer_move_to_active(struct list_head *objects, 889i915_gem_execbuffer_move_to_active(struct list_head *vmas,
813 struct i915_address_space *vm,
814 struct intel_ring_buffer *ring) 890 struct intel_ring_buffer *ring)
815{ 891{
816 struct drm_i915_gem_object *obj; 892 struct i915_vma *vma;
817 893
818 list_for_each_entry(obj, objects, exec_list) { 894 list_for_each_entry(vma, vmas, exec_list) {
895 struct drm_i915_gem_object *obj = vma->obj;
819 u32 old_read = obj->base.read_domains; 896 u32 old_read = obj->base.read_domains;
820 u32 old_write = obj->base.write_domain; 897 u32 old_write = obj->base.write_domain;
821 898
@@ -825,9 +902,7 @@ i915_gem_execbuffer_move_to_active(struct list_head *objects,
825 obj->base.read_domains = obj->base.pending_read_domains; 902 obj->base.read_domains = obj->base.pending_read_domains;
826 obj->fenced_gpu_access = obj->pending_fenced_gpu_access; 903 obj->fenced_gpu_access = obj->pending_fenced_gpu_access;
827 904
828 /* FIXME: This lookup gets fixed later <-- danvet */ 905 i915_vma_move_to_active(vma, ring);
829 list_move_tail(&i915_gem_obj_to_vma(obj, vm)->mm_list, &vm->active_list);
830 i915_gem_object_move_to_active(obj, ring);
831 if (obj->base.write_domain) { 906 if (obj->base.write_domain) {
832 obj->dirty = 1; 907 obj->dirty = 1;
833 obj->last_write_seqno = intel_ring_get_seqno(ring); 908 obj->last_write_seqno = intel_ring_get_seqno(ring);
@@ -885,10 +960,11 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
885 struct i915_address_space *vm) 960 struct i915_address_space *vm)
886{ 961{
887 drm_i915_private_t *dev_priv = dev->dev_private; 962 drm_i915_private_t *dev_priv = dev->dev_private;
888 struct eb_objects *eb; 963 struct eb_vmas *eb;
889 struct drm_i915_gem_object *batch_obj; 964 struct drm_i915_gem_object *batch_obj;
890 struct drm_clip_rect *cliprects = NULL; 965 struct drm_clip_rect *cliprects = NULL;
891 struct intel_ring_buffer *ring; 966 struct intel_ring_buffer *ring;
967 struct i915_ctx_hang_stats *hs;
892 u32 ctx_id = i915_execbuffer2_get_context_id(*args); 968 u32 ctx_id = i915_execbuffer2_get_context_id(*args);
893 u32 exec_start, exec_len; 969 u32 exec_start, exec_len;
894 u32 mask, flags; 970 u32 mask, flags;
@@ -1000,7 +1076,8 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1000 return -EINVAL; 1076 return -EINVAL;
1001 } 1077 }
1002 1078
1003 cliprects = kmalloc(args->num_cliprects * sizeof(*cliprects), 1079 cliprects = kcalloc(args->num_cliprects,
1080 sizeof(*cliprects),
1004 GFP_KERNEL); 1081 GFP_KERNEL);
1005 if (cliprects == NULL) { 1082 if (cliprects == NULL) {
1006 ret = -ENOMEM; 1083 ret = -ENOMEM;
@@ -1025,7 +1102,7 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1025 goto pre_mutex_err; 1102 goto pre_mutex_err;
1026 } 1103 }
1027 1104
1028 eb = eb_create(args); 1105 eb = eb_create(args, vm);
1029 if (eb == NULL) { 1106 if (eb == NULL) {
1030 mutex_unlock(&dev->struct_mutex); 1107 mutex_unlock(&dev->struct_mutex);
1031 ret = -ENOMEM; 1108 ret = -ENOMEM;
@@ -1033,18 +1110,16 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1033 } 1110 }
1034 1111
1035 /* Look up object handles */ 1112 /* Look up object handles */
1036 ret = eb_lookup_objects(eb, exec, args, file); 1113 ret = eb_lookup_vmas(eb, exec, args, vm, file);
1037 if (ret) 1114 if (ret)
1038 goto err; 1115 goto err;
1039 1116
1040 /* take note of the batch buffer before we might reorder the lists */ 1117 /* take note of the batch buffer before we might reorder the lists */
1041 batch_obj = list_entry(eb->objects.prev, 1118 batch_obj = list_entry(eb->vmas.prev, struct i915_vma, exec_list)->obj;
1042 struct drm_i915_gem_object,
1043 exec_list);
1044 1119
1045 /* Move the objects en-masse into the GTT, evicting if necessary. */ 1120 /* Move the objects en-masse into the GTT, evicting if necessary. */
1046 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0; 1121 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
1047 ret = i915_gem_execbuffer_reserve(ring, &eb->objects, vm, &need_relocs); 1122 ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, &need_relocs);
1048 if (ret) 1123 if (ret)
1049 goto err; 1124 goto err;
1050 1125
@@ -1054,7 +1129,7 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1054 if (ret) { 1129 if (ret) {
1055 if (ret == -EFAULT) { 1130 if (ret == -EFAULT) {
1056 ret = i915_gem_execbuffer_relocate_slow(dev, args, file, ring, 1131 ret = i915_gem_execbuffer_relocate_slow(dev, args, file, ring,
1057 eb, exec, vm); 1132 eb, exec);
1058 BUG_ON(!mutex_is_locked(&dev->struct_mutex)); 1133 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1059 } 1134 }
1060 if (ret) 1135 if (ret)
@@ -1071,15 +1146,25 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1071 1146
1072 /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure 1147 /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1073 * batch" bit. Hence we need to pin secure batches into the global gtt. 1148 * batch" bit. Hence we need to pin secure batches into the global gtt.
1074 * hsw should have this fixed, but let's be paranoid and do it 1149 * hsw should have this fixed, but bdw mucks it up again. */
1075 * unconditionally for now. */
1076 if (flags & I915_DISPATCH_SECURE && !batch_obj->has_global_gtt_mapping) 1150 if (flags & I915_DISPATCH_SECURE && !batch_obj->has_global_gtt_mapping)
1077 i915_gem_gtt_bind_object(batch_obj, batch_obj->cache_level); 1151 i915_gem_gtt_bind_object(batch_obj, batch_obj->cache_level);
1078 1152
1079 ret = i915_gem_execbuffer_move_to_gpu(ring, &eb->objects); 1153 ret = i915_gem_execbuffer_move_to_gpu(ring, &eb->vmas);
1080 if (ret) 1154 if (ret)
1081 goto err; 1155 goto err;
1082 1156
1157 hs = i915_gem_context_get_hang_stats(dev, file, ctx_id);
1158 if (IS_ERR(hs)) {
1159 ret = PTR_ERR(hs);
1160 goto err;
1161 }
1162
1163 if (hs->banned) {
1164 ret = -EIO;
1165 goto err;
1166 }
1167
1083 ret = i915_switch_context(ring, file, ctx_id); 1168 ret = i915_switch_context(ring, file, ctx_id);
1084 if (ret) 1169 if (ret)
1085 goto err; 1170 goto err;
@@ -1131,7 +1216,7 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1131 1216
1132 trace_i915_gem_ring_dispatch(ring, intel_ring_get_seqno(ring), flags); 1217 trace_i915_gem_ring_dispatch(ring, intel_ring_get_seqno(ring), flags);
1133 1218
1134 i915_gem_execbuffer_move_to_active(&eb->objects, vm, ring); 1219 i915_gem_execbuffer_move_to_active(&eb->vmas, ring);
1135 i915_gem_execbuffer_retire_commands(dev, file, ring, batch_obj); 1220 i915_gem_execbuffer_retire_commands(dev, file, ring, batch_obj);
1136 1221
1137err: 1222err:
generated by cgit v1.2.2 (git 2.25.0) at 2025-05-29 10:45:06 -0400