1 /*
2 * SPDX-License-Identifier: MIT
3 *
4 * Copyright © 2008-2012 Intel Corporation
5 */
6
7 #include <linux/errno.h>
8 #include <linux/mutex.h>
9
10 #include <drm/drm_mm.h>
11 #include <drm/i915_drm.h>
12
13 #include "gem/i915_gem_lmem.h"
14 #include "gem/i915_gem_region.h"
15 #include "gt/intel_gt.h"
16 #include "gt/intel_gt_mcr.h"
17 #include "gt/intel_gt_regs.h"
18 #include "gt/intel_region_lmem.h"
19 #include "i915_drv.h"
20 #include "i915_gem_stolen.h"
21 #include "i915_pci.h"
22 #include "i915_reg.h"
23 #include "i915_utils.h"
24 #include "i915_vgpu.h"
25 #include "intel_mchbar_regs.h"
26 #include "intel_pci_config.h"
27
28 /*
29 * The BIOS typically reserves some of the system's memory for the exclusive
30 * use of the integrated graphics. This memory is no longer available for
31 * use by the OS and so the user finds that his system has less memory
32 * available than he put in. We refer to this memory as stolen.
33 *
34 * The BIOS will allocate its framebuffer from the stolen memory. Our
35 * goal is try to reuse that object for our own fbcon which must always
36 * be available for panics. Anything else we can reuse the stolen memory
37 * for is a boon.
38 */
39
i915_gem_stolen_insert_node_in_range(struct drm_i915_private * i915,struct drm_mm_node * node,u64 size,unsigned alignment,u64 start,u64 end)40 int i915_gem_stolen_insert_node_in_range(struct drm_i915_private *i915,
41 struct drm_mm_node *node, u64 size,
42 unsigned alignment, u64 start, u64 end)
43 {
44 int ret;
45
46 if (!drm_mm_initialized(&i915->mm.stolen))
47 return -ENODEV;
48
49 /* WaSkipStolenMemoryFirstPage:bdw+ */
50 if (GRAPHICS_VER(i915) >= 8 && start < 4096)
51 start = 4096;
52
53 mutex_lock(&i915->mm.stolen_lock);
54 ret = drm_mm_insert_node_in_range(&i915->mm.stolen, node,
55 size, alignment, 0,
56 start, end, DRM_MM_INSERT_BEST);
57 mutex_unlock(&i915->mm.stolen_lock);
58
59 return ret;
60 }
61
i915_gem_stolen_insert_node(struct drm_i915_private * i915,struct drm_mm_node * node,u64 size,unsigned alignment)62 int i915_gem_stolen_insert_node(struct drm_i915_private *i915,
63 struct drm_mm_node *node, u64 size,
64 unsigned alignment)
65 {
66 return i915_gem_stolen_insert_node_in_range(i915, node,
67 size, alignment,
68 I915_GEM_STOLEN_BIAS,
69 U64_MAX);
70 }
71
i915_gem_stolen_remove_node(struct drm_i915_private * i915,struct drm_mm_node * node)72 void i915_gem_stolen_remove_node(struct drm_i915_private *i915,
73 struct drm_mm_node *node)
74 {
75 mutex_lock(&i915->mm.stolen_lock);
76 drm_mm_remove_node(node);
77 mutex_unlock(&i915->mm.stolen_lock);
78 }
79
i915_adjust_stolen(struct drm_i915_private * i915,struct resource * dsm)80 static int i915_adjust_stolen(struct drm_i915_private *i915,
81 struct resource *dsm)
82 {
83 struct i915_ggtt *ggtt = to_gt(i915)->ggtt;
84 struct intel_uncore *uncore = ggtt->vm.gt->uncore;
85 struct resource *r;
86
87 if (dsm->start == 0 || dsm->end <= dsm->start)
88 return -EINVAL;
89
90 /*
91 * TODO: We have yet too encounter the case where the GTT wasn't at the
92 * end of stolen. With that assumption we could simplify this.
93 */
94
95 /* Make sure we don't clobber the GTT if it's within stolen memory */
96 if (GRAPHICS_VER(i915) <= 4 &&
97 !IS_G33(i915) && !IS_PINEVIEW(i915) && !IS_G4X(i915)) {
98 struct resource stolen[2] = {*dsm, *dsm};
99 struct resource ggtt_res;
100 resource_size_t ggtt_start;
101
102 ggtt_start = intel_uncore_read(uncore, PGTBL_CTL);
103 if (GRAPHICS_VER(i915) == 4)
104 ggtt_start = (ggtt_start & PGTBL_ADDRESS_LO_MASK) |
105 (ggtt_start & PGTBL_ADDRESS_HI_MASK) << 28;
106 else
107 ggtt_start &= PGTBL_ADDRESS_LO_MASK;
108
109 ggtt_res =
110 (struct resource) DEFINE_RES_MEM(ggtt_start,
111 ggtt_total_entries(ggtt) * 4);
112
113 if (ggtt_res.start >= stolen[0].start && ggtt_res.start < stolen[0].end)
114 stolen[0].end = ggtt_res.start;
115 if (ggtt_res.end > stolen[1].start && ggtt_res.end <= stolen[1].end)
116 stolen[1].start = ggtt_res.end;
117
118 /* Pick the larger of the two chunks */
119 if (resource_size(&stolen[0]) > resource_size(&stolen[1]))
120 *dsm = stolen[0];
121 else
122 *dsm = stolen[1];
123
124 if (stolen[0].start != stolen[1].start ||
125 stolen[0].end != stolen[1].end) {
126 drm_dbg(&i915->drm,
127 "GTT within stolen memory at %pR\n",
128 &ggtt_res);
129 drm_dbg(&i915->drm, "Stolen memory adjusted to %pR\n",
130 dsm);
131 }
132 }
133
134 /*
135 * With stolen lmem, we don't need to check if the address range
136 * overlaps with the non-stolen system memory range, since lmem is local
137 * to the gpu.
138 */
139 if (HAS_LMEM(i915))
140 return 0;
141
142 /*
143 * Verify that nothing else uses this physical address. Stolen
144 * memory should be reserved by the BIOS and hidden from the
145 * kernel. So if the region is already marked as busy, something
146 * is seriously wrong.
147 */
148 r = devm_request_mem_region(i915->drm.dev, dsm->start,
149 resource_size(dsm),
150 "Graphics Stolen Memory");
151 if (r == NULL) {
152 /*
153 * One more attempt but this time requesting region from
154 * start + 1, as we have seen that this resolves the region
155 * conflict with the PCI Bus.
156 * This is a BIOS w/a: Some BIOS wrap stolen in the root
157 * PCI bus, but have an off-by-one error. Hence retry the
158 * reservation starting from 1 instead of 0.
159 * There's also BIOS with off-by-one on the other end.
160 */
161 r = devm_request_mem_region(i915->drm.dev, dsm->start + 1,
162 resource_size(dsm) - 2,
163 "Graphics Stolen Memory");
164 /*
165 * GEN3 firmware likes to smash pci bridges into the stolen
166 * range. Apparently this works.
167 */
168 if (!r && GRAPHICS_VER(i915) != 3) {
169 drm_err(&i915->drm,
170 "conflict detected with stolen region: %pR\n",
171 dsm);
172
173 return -EBUSY;
174 }
175 }
176
177 return 0;
178 }
179
i915_gem_cleanup_stolen(struct drm_i915_private * i915)180 static void i915_gem_cleanup_stolen(struct drm_i915_private *i915)
181 {
182 if (!drm_mm_initialized(&i915->mm.stolen))
183 return;
184
185 drm_mm_takedown(&i915->mm.stolen);
186 }
187
g4x_get_stolen_reserved(struct drm_i915_private * i915,struct intel_uncore * uncore,resource_size_t * base,resource_size_t * size)188 static void g4x_get_stolen_reserved(struct drm_i915_private *i915,
189 struct intel_uncore *uncore,
190 resource_size_t *base,
191 resource_size_t *size)
192 {
193 u32 reg_val = intel_uncore_read(uncore,
194 IS_GM45(i915) ?
195 CTG_STOLEN_RESERVED :
196 ELK_STOLEN_RESERVED);
197 resource_size_t stolen_top = i915->dsm.end + 1;
198
199 drm_dbg(&i915->drm, "%s_STOLEN_RESERVED = %08x\n",
200 IS_GM45(i915) ? "CTG" : "ELK", reg_val);
201
202 if ((reg_val & G4X_STOLEN_RESERVED_ENABLE) == 0)
203 return;
204
205 /*
206 * Whether ILK really reuses the ELK register for this is unclear.
207 * Let's see if we catch anyone with this supposedly enabled on ILK.
208 */
209 drm_WARN(&i915->drm, GRAPHICS_VER(i915) == 5,
210 "ILK stolen reserved found? 0x%08x\n",
211 reg_val);
212
213 if (!(reg_val & G4X_STOLEN_RESERVED_ADDR2_MASK))
214 return;
215
216 *base = (reg_val & G4X_STOLEN_RESERVED_ADDR2_MASK) << 16;
217 drm_WARN_ON(&i915->drm,
218 (reg_val & G4X_STOLEN_RESERVED_ADDR1_MASK) < *base);
219
220 *size = stolen_top - *base;
221 }
222
gen6_get_stolen_reserved(struct drm_i915_private * i915,struct intel_uncore * uncore,resource_size_t * base,resource_size_t * size)223 static void gen6_get_stolen_reserved(struct drm_i915_private *i915,
224 struct intel_uncore *uncore,
225 resource_size_t *base,
226 resource_size_t *size)
227 {
228 u32 reg_val = intel_uncore_read(uncore, GEN6_STOLEN_RESERVED);
229
230 drm_dbg(&i915->drm, "GEN6_STOLEN_RESERVED = %08x\n", reg_val);
231
232 if (!(reg_val & GEN6_STOLEN_RESERVED_ENABLE))
233 return;
234
235 *base = reg_val & GEN6_STOLEN_RESERVED_ADDR_MASK;
236
237 switch (reg_val & GEN6_STOLEN_RESERVED_SIZE_MASK) {
238 case GEN6_STOLEN_RESERVED_1M:
239 *size = 1024 * 1024;
240 break;
241 case GEN6_STOLEN_RESERVED_512K:
242 *size = 512 * 1024;
243 break;
244 case GEN6_STOLEN_RESERVED_256K:
245 *size = 256 * 1024;
246 break;
247 case GEN6_STOLEN_RESERVED_128K:
248 *size = 128 * 1024;
249 break;
250 default:
251 *size = 1024 * 1024;
252 MISSING_CASE(reg_val & GEN6_STOLEN_RESERVED_SIZE_MASK);
253 }
254 }
255
vlv_get_stolen_reserved(struct drm_i915_private * i915,struct intel_uncore * uncore,resource_size_t * base,resource_size_t * size)256 static void vlv_get_stolen_reserved(struct drm_i915_private *i915,
257 struct intel_uncore *uncore,
258 resource_size_t *base,
259 resource_size_t *size)
260 {
261 u32 reg_val = intel_uncore_read(uncore, GEN6_STOLEN_RESERVED);
262 resource_size_t stolen_top = i915->dsm.end + 1;
263
264 drm_dbg(&i915->drm, "GEN6_STOLEN_RESERVED = %08x\n", reg_val);
265
266 if (!(reg_val & GEN6_STOLEN_RESERVED_ENABLE))
267 return;
268
269 switch (reg_val & GEN7_STOLEN_RESERVED_SIZE_MASK) {
270 default:
271 MISSING_CASE(reg_val & GEN7_STOLEN_RESERVED_SIZE_MASK);
272 fallthrough;
273 case GEN7_STOLEN_RESERVED_1M:
274 *size = 1024 * 1024;
275 break;
276 }
277
278 /*
279 * On vlv, the ADDR_MASK portion is left as 0 and HW deduces the
280 * reserved location as (top - size).
281 */
282 *base = stolen_top - *size;
283 }
284
gen7_get_stolen_reserved(struct drm_i915_private * i915,struct intel_uncore * uncore,resource_size_t * base,resource_size_t * size)285 static void gen7_get_stolen_reserved(struct drm_i915_private *i915,
286 struct intel_uncore *uncore,
287 resource_size_t *base,
288 resource_size_t *size)
289 {
290 u32 reg_val = intel_uncore_read(uncore, GEN6_STOLEN_RESERVED);
291
292 drm_dbg(&i915->drm, "GEN6_STOLEN_RESERVED = %08x\n", reg_val);
293
294 if (!(reg_val & GEN6_STOLEN_RESERVED_ENABLE))
295 return;
296
297 *base = reg_val & GEN7_STOLEN_RESERVED_ADDR_MASK;
298
299 switch (reg_val & GEN7_STOLEN_RESERVED_SIZE_MASK) {
300 case GEN7_STOLEN_RESERVED_1M:
301 *size = 1024 * 1024;
302 break;
303 case GEN7_STOLEN_RESERVED_256K:
304 *size = 256 * 1024;
305 break;
306 default:
307 *size = 1024 * 1024;
308 MISSING_CASE(reg_val & GEN7_STOLEN_RESERVED_SIZE_MASK);
309 }
310 }
311
chv_get_stolen_reserved(struct drm_i915_private * i915,struct intel_uncore * uncore,resource_size_t * base,resource_size_t * size)312 static void chv_get_stolen_reserved(struct drm_i915_private *i915,
313 struct intel_uncore *uncore,
314 resource_size_t *base,
315 resource_size_t *size)
316 {
317 u32 reg_val = intel_uncore_read(uncore, GEN6_STOLEN_RESERVED);
318
319 drm_dbg(&i915->drm, "GEN6_STOLEN_RESERVED = %08x\n", reg_val);
320
321 if (!(reg_val & GEN6_STOLEN_RESERVED_ENABLE))
322 return;
323
324 *base = reg_val & GEN6_STOLEN_RESERVED_ADDR_MASK;
325
326 switch (reg_val & GEN8_STOLEN_RESERVED_SIZE_MASK) {
327 case GEN8_STOLEN_RESERVED_1M:
328 *size = 1024 * 1024;
329 break;
330 case GEN8_STOLEN_RESERVED_2M:
331 *size = 2 * 1024 * 1024;
332 break;
333 case GEN8_STOLEN_RESERVED_4M:
334 *size = 4 * 1024 * 1024;
335 break;
336 case GEN8_STOLEN_RESERVED_8M:
337 *size = 8 * 1024 * 1024;
338 break;
339 default:
340 *size = 8 * 1024 * 1024;
341 MISSING_CASE(reg_val & GEN8_STOLEN_RESERVED_SIZE_MASK);
342 }
343 }
344
bdw_get_stolen_reserved(struct drm_i915_private * i915,struct intel_uncore * uncore,resource_size_t * base,resource_size_t * size)345 static void bdw_get_stolen_reserved(struct drm_i915_private *i915,
346 struct intel_uncore *uncore,
347 resource_size_t *base,
348 resource_size_t *size)
349 {
350 u32 reg_val = intel_uncore_read(uncore, GEN6_STOLEN_RESERVED);
351 resource_size_t stolen_top = i915->dsm.end + 1;
352
353 drm_dbg(&i915->drm, "GEN6_STOLEN_RESERVED = %08x\n", reg_val);
354
355 if (!(reg_val & GEN6_STOLEN_RESERVED_ENABLE))
356 return;
357
358 if (!(reg_val & GEN6_STOLEN_RESERVED_ADDR_MASK))
359 return;
360
361 *base = reg_val & GEN6_STOLEN_RESERVED_ADDR_MASK;
362 *size = stolen_top - *base;
363 }
364
icl_get_stolen_reserved(struct drm_i915_private * i915,struct intel_uncore * uncore,resource_size_t * base,resource_size_t * size)365 static void icl_get_stolen_reserved(struct drm_i915_private *i915,
366 struct intel_uncore *uncore,
367 resource_size_t *base,
368 resource_size_t *size)
369 {
370 u64 reg_val = intel_uncore_read64(uncore, GEN6_STOLEN_RESERVED);
371
372 drm_dbg(&i915->drm, "GEN6_STOLEN_RESERVED = 0x%016llx\n", reg_val);
373
374 *base = reg_val & GEN11_STOLEN_RESERVED_ADDR_MASK;
375
376 switch (reg_val & GEN8_STOLEN_RESERVED_SIZE_MASK) {
377 case GEN8_STOLEN_RESERVED_1M:
378 *size = 1024 * 1024;
379 break;
380 case GEN8_STOLEN_RESERVED_2M:
381 *size = 2 * 1024 * 1024;
382 break;
383 case GEN8_STOLEN_RESERVED_4M:
384 *size = 4 * 1024 * 1024;
385 break;
386 case GEN8_STOLEN_RESERVED_8M:
387 *size = 8 * 1024 * 1024;
388 break;
389 default:
390 *size = 8 * 1024 * 1024;
391 MISSING_CASE(reg_val & GEN8_STOLEN_RESERVED_SIZE_MASK);
392 }
393 }
394
i915_gem_init_stolen(struct intel_memory_region * mem)395 static int i915_gem_init_stolen(struct intel_memory_region *mem)
396 {
397 struct drm_i915_private *i915 = mem->i915;
398 struct intel_uncore *uncore = &i915->uncore;
399 resource_size_t reserved_base, stolen_top;
400 resource_size_t reserved_total, reserved_size;
401
402 mutex_init(&i915->mm.stolen_lock);
403
404 if (intel_vgpu_active(i915)) {
405 drm_notice(&i915->drm,
406 "%s, disabling use of stolen memory\n",
407 "iGVT-g active");
408 return 0;
409 }
410
411 if (i915_vtd_active(i915) && GRAPHICS_VER(i915) < 8) {
412 drm_notice(&i915->drm,
413 "%s, disabling use of stolen memory\n",
414 "DMAR active");
415 return 0;
416 }
417
418 if (resource_size(&mem->region) == 0)
419 return 0;
420
421 i915->dsm = mem->region;
422
423 if (i915_adjust_stolen(i915, &i915->dsm))
424 return 0;
425
426 GEM_BUG_ON(i915->dsm.start == 0);
427 GEM_BUG_ON(i915->dsm.end <= i915->dsm.start);
428
429 stolen_top = i915->dsm.end + 1;
430 reserved_base = stolen_top;
431 reserved_size = 0;
432
433 if (GRAPHICS_VER(i915) >= 11) {
434 icl_get_stolen_reserved(i915, uncore,
435 &reserved_base, &reserved_size);
436 } else if (GRAPHICS_VER(i915) >= 8) {
437 if (IS_LP(i915))
438 chv_get_stolen_reserved(i915, uncore,
439 &reserved_base, &reserved_size);
440 else
441 bdw_get_stolen_reserved(i915, uncore,
442 &reserved_base, &reserved_size);
443 } else if (GRAPHICS_VER(i915) >= 7) {
444 if (IS_VALLEYVIEW(i915))
445 vlv_get_stolen_reserved(i915, uncore,
446 &reserved_base, &reserved_size);
447 else
448 gen7_get_stolen_reserved(i915, uncore,
449 &reserved_base, &reserved_size);
450 } else if (GRAPHICS_VER(i915) >= 6) {
451 gen6_get_stolen_reserved(i915, uncore,
452 &reserved_base, &reserved_size);
453 } else if (GRAPHICS_VER(i915) >= 5 || IS_G4X(i915)) {
454 g4x_get_stolen_reserved(i915, uncore,
455 &reserved_base, &reserved_size);
456 }
457
458 /*
459 * Our expectation is that the reserved space is at the top of the
460 * stolen region and *never* at the bottom. If we see !reserved_base,
461 * it likely means we failed to read the registers correctly.
462 */
463 if (!reserved_base) {
464 drm_err(&i915->drm,
465 "inconsistent reservation %pa + %pa; ignoring\n",
466 &reserved_base, &reserved_size);
467 reserved_base = stolen_top;
468 reserved_size = 0;
469 }
470
471 i915->dsm_reserved =
472 (struct resource)DEFINE_RES_MEM(reserved_base, reserved_size);
473
474 if (!resource_contains(&i915->dsm, &i915->dsm_reserved)) {
475 drm_err(&i915->drm,
476 "Stolen reserved area %pR outside stolen memory %pR\n",
477 &i915->dsm_reserved, &i915->dsm);
478 return 0;
479 }
480
481 /* Exclude the reserved region from driver use */
482 mem->region.end = reserved_base - 1;
483 mem->io_size = min(mem->io_size, resource_size(&mem->region));
484
485 /* It is possible for the reserved area to end before the end of stolen
486 * memory, so just consider the start. */
487 reserved_total = stolen_top - reserved_base;
488
489 i915->stolen_usable_size =
490 resource_size(&i915->dsm) - reserved_total;
491
492 drm_dbg(&i915->drm,
493 "Memory reserved for graphics device: %lluK, usable: %lluK\n",
494 (u64)resource_size(&i915->dsm) >> 10,
495 (u64)i915->stolen_usable_size >> 10);
496
497 if (i915->stolen_usable_size == 0)
498 return 0;
499
500 /* Basic memrange allocator for stolen space. */
501 drm_mm_init(&i915->mm.stolen, 0, i915->stolen_usable_size);
502
503 return 0;
504 }
505
dbg_poison(struct i915_ggtt * ggtt,dma_addr_t addr,resource_size_t size,u8 x)506 static void dbg_poison(struct i915_ggtt *ggtt,
507 dma_addr_t addr, resource_size_t size,
508 u8 x)
509 {
510 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)
511 if (!drm_mm_node_allocated(&ggtt->error_capture))
512 return;
513
514 if (ggtt->vm.bind_async_flags & I915_VMA_GLOBAL_BIND)
515 return; /* beware stop_machine() inversion */
516
517 GEM_BUG_ON(!IS_ALIGNED(size, PAGE_SIZE));
518
519 mutex_lock(&ggtt->error_mutex);
520 while (size) {
521 void __iomem *s;
522
523 ggtt->vm.insert_page(&ggtt->vm, addr,
524 ggtt->error_capture.start,
525 I915_CACHE_NONE, 0);
526 mb();
527
528 s = io_mapping_map_wc(&ggtt->iomap,
529 ggtt->error_capture.start,
530 PAGE_SIZE);
531 memset_io(s, x, PAGE_SIZE);
532 io_mapping_unmap(s);
533
534 addr += PAGE_SIZE;
535 size -= PAGE_SIZE;
536 }
537 mb();
538 ggtt->vm.clear_range(&ggtt->vm, ggtt->error_capture.start, PAGE_SIZE);
539 mutex_unlock(&ggtt->error_mutex);
540 #endif
541 }
542
543 static struct sg_table *
i915_pages_create_for_stolen(struct drm_device * dev,resource_size_t offset,resource_size_t size)544 i915_pages_create_for_stolen(struct drm_device *dev,
545 resource_size_t offset, resource_size_t size)
546 {
547 struct drm_i915_private *i915 = to_i915(dev);
548 struct sg_table *st;
549 struct scatterlist *sg;
550
551 GEM_BUG_ON(range_overflows(offset, size, resource_size(&i915->dsm)));
552
553 /* We hide that we have no struct page backing our stolen object
554 * by wrapping the contiguous physical allocation with a fake
555 * dma mapping in a single scatterlist.
556 */
557
558 st = kmalloc(sizeof(*st), GFP_KERNEL);
559 if (st == NULL)
560 return ERR_PTR(-ENOMEM);
561
562 if (sg_alloc_table(st, 1, GFP_KERNEL)) {
563 kfree(st);
564 return ERR_PTR(-ENOMEM);
565 }
566
567 sg = st->sgl;
568 sg->offset = 0;
569 sg->length = size;
570
571 sg_dma_address(sg) = (dma_addr_t)i915->dsm.start + offset;
572 sg_dma_len(sg) = size;
573
574 return st;
575 }
576
i915_gem_object_get_pages_stolen(struct drm_i915_gem_object * obj)577 static int i915_gem_object_get_pages_stolen(struct drm_i915_gem_object *obj)
578 {
579 struct drm_i915_private *i915 = to_i915(obj->base.dev);
580 struct sg_table *pages =
581 i915_pages_create_for_stolen(obj->base.dev,
582 obj->stolen->start,
583 obj->stolen->size);
584 if (IS_ERR(pages))
585 return PTR_ERR(pages);
586
587 dbg_poison(to_gt(i915)->ggtt,
588 sg_dma_address(pages->sgl),
589 sg_dma_len(pages->sgl),
590 POISON_INUSE);
591
592 __i915_gem_object_set_pages(obj, pages, obj->stolen->size);
593
594 return 0;
595 }
596
i915_gem_object_put_pages_stolen(struct drm_i915_gem_object * obj,struct sg_table * pages)597 static void i915_gem_object_put_pages_stolen(struct drm_i915_gem_object *obj,
598 struct sg_table *pages)
599 {
600 struct drm_i915_private *i915 = to_i915(obj->base.dev);
601 /* Should only be called from i915_gem_object_release_stolen() */
602
603 dbg_poison(to_gt(i915)->ggtt,
604 sg_dma_address(pages->sgl),
605 sg_dma_len(pages->sgl),
606 POISON_FREE);
607
608 sg_free_table(pages);
609 kfree(pages);
610 }
611
612 static void
i915_gem_object_release_stolen(struct drm_i915_gem_object * obj)613 i915_gem_object_release_stolen(struct drm_i915_gem_object *obj)
614 {
615 struct drm_i915_private *i915 = to_i915(obj->base.dev);
616 struct drm_mm_node *stolen = fetch_and_zero(&obj->stolen);
617
618 GEM_BUG_ON(!stolen);
619 i915_gem_stolen_remove_node(i915, stolen);
620 kfree(stolen);
621
622 i915_gem_object_release_memory_region(obj);
623 }
624
625 static const struct drm_i915_gem_object_ops i915_gem_object_stolen_ops = {
626 .name = "i915_gem_object_stolen",
627 .get_pages = i915_gem_object_get_pages_stolen,
628 .put_pages = i915_gem_object_put_pages_stolen,
629 .release = i915_gem_object_release_stolen,
630 };
631
__i915_gem_object_create_stolen(struct intel_memory_region * mem,struct drm_i915_gem_object * obj,struct drm_mm_node * stolen)632 static int __i915_gem_object_create_stolen(struct intel_memory_region *mem,
633 struct drm_i915_gem_object *obj,
634 struct drm_mm_node *stolen)
635 {
636 static struct lock_class_key lock_class;
637 unsigned int cache_level;
638 unsigned int flags;
639 int err;
640
641 /*
642 * Stolen objects are always physically contiguous since we just
643 * allocate one big block underneath using the drm_mm range allocator.
644 */
645 flags = I915_BO_ALLOC_CONTIGUOUS;
646
647 drm_gem_private_object_init(&mem->i915->drm, &obj->base, stolen->size);
648 i915_gem_object_init(obj, &i915_gem_object_stolen_ops, &lock_class, flags);
649
650 obj->stolen = stolen;
651 obj->read_domains = I915_GEM_DOMAIN_CPU | I915_GEM_DOMAIN_GTT;
652 cache_level = HAS_LLC(mem->i915) ? I915_CACHE_LLC : I915_CACHE_NONE;
653 i915_gem_object_set_cache_coherency(obj, cache_level);
654
655 if (WARN_ON(!i915_gem_object_trylock(obj, NULL)))
656 return -EBUSY;
657
658 i915_gem_object_init_memory_region(obj, mem);
659
660 err = i915_gem_object_pin_pages(obj);
661 if (err)
662 i915_gem_object_release_memory_region(obj);
663 i915_gem_object_unlock(obj);
664
665 return err;
666 }
667
_i915_gem_object_stolen_init(struct intel_memory_region * mem,struct drm_i915_gem_object * obj,resource_size_t offset,resource_size_t size,resource_size_t page_size,unsigned int flags)668 static int _i915_gem_object_stolen_init(struct intel_memory_region *mem,
669 struct drm_i915_gem_object *obj,
670 resource_size_t offset,
671 resource_size_t size,
672 resource_size_t page_size,
673 unsigned int flags)
674 {
675 struct drm_i915_private *i915 = mem->i915;
676 struct drm_mm_node *stolen;
677 int ret;
678
679 if (!drm_mm_initialized(&i915->mm.stolen))
680 return -ENODEV;
681
682 if (size == 0)
683 return -EINVAL;
684
685 /*
686 * With discrete devices, where we lack a mappable aperture there is no
687 * possible way to ever access this memory on the CPU side.
688 */
689 if (mem->type == INTEL_MEMORY_STOLEN_LOCAL && !mem->io_size &&
690 !(flags & I915_BO_ALLOC_GPU_ONLY))
691 return -ENOSPC;
692
693 stolen = kzalloc(sizeof(*stolen), GFP_KERNEL);
694 if (!stolen)
695 return -ENOMEM;
696
697 if (offset != I915_BO_INVALID_OFFSET) {
698 drm_dbg(&i915->drm,
699 "creating preallocated stolen object: stolen_offset=%pa, size=%pa\n",
700 &offset, &size);
701
702 stolen->start = offset;
703 stolen->size = size;
704 mutex_lock(&i915->mm.stolen_lock);
705 ret = drm_mm_reserve_node(&i915->mm.stolen, stolen);
706 mutex_unlock(&i915->mm.stolen_lock);
707 } else {
708 ret = i915_gem_stolen_insert_node(i915, stolen, size,
709 mem->min_page_size);
710 }
711 if (ret)
712 goto err_free;
713
714 ret = __i915_gem_object_create_stolen(mem, obj, stolen);
715 if (ret)
716 goto err_remove;
717
718 return 0;
719
720 err_remove:
721 i915_gem_stolen_remove_node(i915, stolen);
722 err_free:
723 kfree(stolen);
724 return ret;
725 }
726
727 struct drm_i915_gem_object *
i915_gem_object_create_stolen(struct drm_i915_private * i915,resource_size_t size)728 i915_gem_object_create_stolen(struct drm_i915_private *i915,
729 resource_size_t size)
730 {
731 return i915_gem_object_create_region(i915->mm.stolen_region, size, 0, 0);
732 }
733
init_stolen_smem(struct intel_memory_region * mem)734 static int init_stolen_smem(struct intel_memory_region *mem)
735 {
736 /*
737 * Initialise stolen early so that we may reserve preallocated
738 * objects for the BIOS to KMS transition.
739 */
740 return i915_gem_init_stolen(mem);
741 }
742
release_stolen_smem(struct intel_memory_region * mem)743 static int release_stolen_smem(struct intel_memory_region *mem)
744 {
745 i915_gem_cleanup_stolen(mem->i915);
746 return 0;
747 }
748
749 static const struct intel_memory_region_ops i915_region_stolen_smem_ops = {
750 .init = init_stolen_smem,
751 .release = release_stolen_smem,
752 .init_object = _i915_gem_object_stolen_init,
753 };
754
init_stolen_lmem(struct intel_memory_region * mem)755 static int init_stolen_lmem(struct intel_memory_region *mem)
756 {
757 int err;
758
759 if (GEM_WARN_ON(resource_size(&mem->region) == 0))
760 return -ENODEV;
761
762 /*
763 * TODO: For stolen lmem we mostly just care about populating the dsm
764 * related bits and setting up the drm_mm allocator for the range.
765 * Perhaps split up i915_gem_init_stolen() for this.
766 */
767 err = i915_gem_init_stolen(mem);
768 if (err)
769 return err;
770
771 if (mem->io_size && !io_mapping_init_wc(&mem->iomap,
772 mem->io_start,
773 mem->io_size)) {
774 err = -EIO;
775 goto err_cleanup;
776 }
777
778 return 0;
779
780 err_cleanup:
781 i915_gem_cleanup_stolen(mem->i915);
782 return err;
783 }
784
release_stolen_lmem(struct intel_memory_region * mem)785 static int release_stolen_lmem(struct intel_memory_region *mem)
786 {
787 if (mem->io_size)
788 io_mapping_fini(&mem->iomap);
789 i915_gem_cleanup_stolen(mem->i915);
790 return 0;
791 }
792
793 static const struct intel_memory_region_ops i915_region_stolen_lmem_ops = {
794 .init = init_stolen_lmem,
795 .release = release_stolen_lmem,
796 .init_object = _i915_gem_object_stolen_init,
797 };
798
799 struct intel_memory_region *
i915_gem_stolen_lmem_setup(struct drm_i915_private * i915,u16 type,u16 instance)800 i915_gem_stolen_lmem_setup(struct drm_i915_private *i915, u16 type,
801 u16 instance)
802 {
803 struct intel_uncore *uncore = &i915->uncore;
804 struct pci_dev *pdev = to_pci_dev(i915->drm.dev);
805 resource_size_t dsm_size, dsm_base, lmem_size;
806 struct intel_memory_region *mem;
807 resource_size_t io_start, io_size;
808 resource_size_t min_page_size;
809
810 if (WARN_ON_ONCE(instance))
811 return ERR_PTR(-ENODEV);
812
813 if (!i915_pci_resource_valid(pdev, GEN12_LMEM_BAR))
814 return ERR_PTR(-ENXIO);
815
816 /* Use DSM base address instead for stolen memory */
817 dsm_base = intel_uncore_read64(uncore, GEN12_DSMBASE);
818 if (IS_DG1(uncore->i915)) {
819 lmem_size = pci_resource_len(pdev, GEN12_LMEM_BAR);
820 if (WARN_ON(lmem_size < dsm_base))
821 return ERR_PTR(-ENODEV);
822 } else {
823 resource_size_t lmem_range;
824
825 lmem_range = intel_gt_mcr_read_any(&i915->gt0, XEHP_TILE0_ADDR_RANGE) & 0xFFFF;
826 lmem_size = lmem_range >> XEHP_TILE_LMEM_RANGE_SHIFT;
827 lmem_size *= SZ_1G;
828 }
829
830 dsm_size = lmem_size - dsm_base;
831 if (pci_resource_len(pdev, GEN12_LMEM_BAR) < lmem_size) {
832 io_start = 0;
833 io_size = 0;
834 } else {
835 io_start = pci_resource_start(pdev, GEN12_LMEM_BAR) + dsm_base;
836 io_size = dsm_size;
837 }
838
839 min_page_size = HAS_64K_PAGES(i915) ? I915_GTT_PAGE_SIZE_64K :
840 I915_GTT_PAGE_SIZE_4K;
841
842 mem = intel_memory_region_create(i915, dsm_base, dsm_size,
843 min_page_size,
844 io_start, io_size,
845 type, instance,
846 &i915_region_stolen_lmem_ops);
847 if (IS_ERR(mem))
848 return mem;
849
850 /*
851 * TODO: consider creating common helper to just print all the
852 * interesting stuff from intel_memory_region, which we can use for all
853 * our probed regions.
854 */
855
856 drm_dbg(&i915->drm, "Stolen Local memory IO start: %pa\n",
857 &mem->io_start);
858 drm_dbg(&i915->drm, "Stolen Local DSM base: %pa\n", &dsm_base);
859
860 intel_memory_region_set_name(mem, "stolen-local");
861
862 mem->private = true;
863
864 return mem;
865 }
866
867 struct intel_memory_region*
i915_gem_stolen_smem_setup(struct drm_i915_private * i915,u16 type,u16 instance)868 i915_gem_stolen_smem_setup(struct drm_i915_private *i915, u16 type,
869 u16 instance)
870 {
871 struct intel_memory_region *mem;
872
873 mem = intel_memory_region_create(i915,
874 intel_graphics_stolen_res.start,
875 resource_size(&intel_graphics_stolen_res),
876 PAGE_SIZE, 0, 0, type, instance,
877 &i915_region_stolen_smem_ops);
878 if (IS_ERR(mem))
879 return mem;
880
881 intel_memory_region_set_name(mem, "stolen-system");
882
883 mem->private = true;
884 return mem;
885 }
886
i915_gem_object_is_stolen(const struct drm_i915_gem_object * obj)887 bool i915_gem_object_is_stolen(const struct drm_i915_gem_object *obj)
888 {
889 return obj->ops == &i915_gem_object_stolen_ops;
890 }
891