Merge tag 'iommu-fixes-v5.15-rc0' of git://git.kernel.org/pub/scm/linux/kernel/git...
[platform/kernel/linux-starfive.git] / drivers / gpu / drm / i915 / gt / gen8_ppgtt.c
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2020 Intel Corporation
4  */
5
6 #include <linux/log2.h>
7
8 #include "gem/i915_gem_lmem.h"
9
10 #include "gen8_ppgtt.h"
11 #include "i915_scatterlist.h"
12 #include "i915_trace.h"
13 #include "i915_pvinfo.h"
14 #include "i915_vgpu.h"
15 #include "intel_gt.h"
16 #include "intel_gtt.h"
17
18 static u64 gen8_pde_encode(const dma_addr_t addr,
19                            const enum i915_cache_level level)
20 {
21         u64 pde = addr | _PAGE_PRESENT | _PAGE_RW;
22
23         if (level != I915_CACHE_NONE)
24                 pde |= PPAT_CACHED_PDE;
25         else
26                 pde |= PPAT_UNCACHED;
27
28         return pde;
29 }
30
31 static u64 gen8_pte_encode(dma_addr_t addr,
32                            enum i915_cache_level level,
33                            u32 flags)
34 {
35         gen8_pte_t pte = addr | _PAGE_PRESENT | _PAGE_RW;
36
37         if (unlikely(flags & PTE_READ_ONLY))
38                 pte &= ~_PAGE_RW;
39
40         if (flags & PTE_LM)
41                 pte |= GEN12_PPGTT_PTE_LM;
42
43         switch (level) {
44         case I915_CACHE_NONE:
45                 pte |= PPAT_UNCACHED;
46                 break;
47         case I915_CACHE_WT:
48                 pte |= PPAT_DISPLAY_ELLC;
49                 break;
50         default:
51                 pte |= PPAT_CACHED;
52                 break;
53         }
54
55         return pte;
56 }
57
58 static void gen8_ppgtt_notify_vgt(struct i915_ppgtt *ppgtt, bool create)
59 {
60         struct drm_i915_private *i915 = ppgtt->vm.i915;
61         struct intel_uncore *uncore = ppgtt->vm.gt->uncore;
62         enum vgt_g2v_type msg;
63         int i;
64
65         if (create)
66                 atomic_inc(px_used(ppgtt->pd)); /* never remove */
67         else
68                 atomic_dec(px_used(ppgtt->pd));
69
70         mutex_lock(&i915->vgpu.lock);
71
72         if (i915_vm_is_4lvl(&ppgtt->vm)) {
73                 const u64 daddr = px_dma(ppgtt->pd);
74
75                 intel_uncore_write(uncore,
76                                    vgtif_reg(pdp[0].lo), lower_32_bits(daddr));
77                 intel_uncore_write(uncore,
78                                    vgtif_reg(pdp[0].hi), upper_32_bits(daddr));
79
80                 msg = create ?
81                         VGT_G2V_PPGTT_L4_PAGE_TABLE_CREATE :
82                         VGT_G2V_PPGTT_L4_PAGE_TABLE_DESTROY;
83         } else {
84                 for (i = 0; i < GEN8_3LVL_PDPES; i++) {
85                         const u64 daddr = i915_page_dir_dma_addr(ppgtt, i);
86
87                         intel_uncore_write(uncore,
88                                            vgtif_reg(pdp[i].lo),
89                                            lower_32_bits(daddr));
90                         intel_uncore_write(uncore,
91                                            vgtif_reg(pdp[i].hi),
92                                            upper_32_bits(daddr));
93                 }
94
95                 msg = create ?
96                         VGT_G2V_PPGTT_L3_PAGE_TABLE_CREATE :
97                         VGT_G2V_PPGTT_L3_PAGE_TABLE_DESTROY;
98         }
99
100         /* g2v_notify atomically (via hv trap) consumes the message packet. */
101         intel_uncore_write(uncore, vgtif_reg(g2v_notify), msg);
102
103         mutex_unlock(&i915->vgpu.lock);
104 }
105
106 /* Index shifts into the pagetable are offset by GEN8_PTE_SHIFT [12] */
107 #define GEN8_PAGE_SIZE (SZ_4K) /* page and page-directory sizes are the same */
108 #define GEN8_PTE_SHIFT (ilog2(GEN8_PAGE_SIZE))
109 #define GEN8_PDES (GEN8_PAGE_SIZE / sizeof(u64))
110 #define gen8_pd_shift(lvl) ((lvl) * ilog2(GEN8_PDES))
111 #define gen8_pd_index(i, lvl) i915_pde_index((i), gen8_pd_shift(lvl))
112 #define __gen8_pte_shift(lvl) (GEN8_PTE_SHIFT + gen8_pd_shift(lvl))
113 #define __gen8_pte_index(a, lvl) i915_pde_index((a), __gen8_pte_shift(lvl))
114
115 #define as_pd(x) container_of((x), typeof(struct i915_page_directory), pt)
116
117 static unsigned int
118 gen8_pd_range(u64 start, u64 end, int lvl, unsigned int *idx)
119 {
120         const int shift = gen8_pd_shift(lvl);
121         const u64 mask = ~0ull << gen8_pd_shift(lvl + 1);
122
123         GEM_BUG_ON(start >= end);
124         end += ~mask >> gen8_pd_shift(1);
125
126         *idx = i915_pde_index(start, shift);
127         if ((start ^ end) & mask)
128                 return GEN8_PDES - *idx;
129         else
130                 return i915_pde_index(end, shift) - *idx;
131 }
132
133 static bool gen8_pd_contains(u64 start, u64 end, int lvl)
134 {
135         const u64 mask = ~0ull << gen8_pd_shift(lvl + 1);
136
137         GEM_BUG_ON(start >= end);
138         return (start ^ end) & mask && (start & ~mask) == 0;
139 }
140
141 static unsigned int gen8_pt_count(u64 start, u64 end)
142 {
143         GEM_BUG_ON(start >= end);
144         if ((start ^ end) >> gen8_pd_shift(1))
145                 return GEN8_PDES - (start & (GEN8_PDES - 1));
146         else
147                 return end - start;
148 }
149
150 static unsigned int gen8_pd_top_count(const struct i915_address_space *vm)
151 {
152         unsigned int shift = __gen8_pte_shift(vm->top);
153
154         return (vm->total + (1ull << shift) - 1) >> shift;
155 }
156
157 static struct i915_page_directory *
158 gen8_pdp_for_page_index(struct i915_address_space * const vm, const u64 idx)
159 {
160         struct i915_ppgtt * const ppgtt = i915_vm_to_ppgtt(vm);
161
162         if (vm->top == 2)
163                 return ppgtt->pd;
164         else
165                 return i915_pd_entry(ppgtt->pd, gen8_pd_index(idx, vm->top));
166 }
167
168 static struct i915_page_directory *
169 gen8_pdp_for_page_address(struct i915_address_space * const vm, const u64 addr)
170 {
171         return gen8_pdp_for_page_index(vm, addr >> GEN8_PTE_SHIFT);
172 }
173
174 static void __gen8_ppgtt_cleanup(struct i915_address_space *vm,
175                                  struct i915_page_directory *pd,
176                                  int count, int lvl)
177 {
178         if (lvl) {
179                 void **pde = pd->entry;
180
181                 do {
182                         if (!*pde)
183                                 continue;
184
185                         __gen8_ppgtt_cleanup(vm, *pde, GEN8_PDES, lvl - 1);
186                 } while (pde++, --count);
187         }
188
189         free_px(vm, &pd->pt, lvl);
190 }
191
192 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
193 {
194         struct i915_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
195
196         if (intel_vgpu_active(vm->i915))
197                 gen8_ppgtt_notify_vgt(ppgtt, false);
198
199         __gen8_ppgtt_cleanup(vm, ppgtt->pd, gen8_pd_top_count(vm), vm->top);
200         free_scratch(vm);
201 }
202
203 static u64 __gen8_ppgtt_clear(struct i915_address_space * const vm,
204                               struct i915_page_directory * const pd,
205                               u64 start, const u64 end, int lvl)
206 {
207         const struct drm_i915_gem_object * const scratch = vm->scratch[lvl];
208         unsigned int idx, len;
209
210         GEM_BUG_ON(end > vm->total >> GEN8_PTE_SHIFT);
211
212         len = gen8_pd_range(start, end, lvl--, &idx);
213         DBG("%s(%p):{ lvl:%d, start:%llx, end:%llx, idx:%d, len:%d, used:%d }\n",
214             __func__, vm, lvl + 1, start, end,
215             idx, len, atomic_read(px_used(pd)));
216         GEM_BUG_ON(!len || len >= atomic_read(px_used(pd)));
217
218         do {
219                 struct i915_page_table *pt = pd->entry[idx];
220
221                 if (atomic_fetch_inc(&pt->used) >> gen8_pd_shift(1) &&
222                     gen8_pd_contains(start, end, lvl)) {
223                         DBG("%s(%p):{ lvl:%d, idx:%d, start:%llx, end:%llx } removing pd\n",
224                             __func__, vm, lvl + 1, idx, start, end);
225                         clear_pd_entry(pd, idx, scratch);
226                         __gen8_ppgtt_cleanup(vm, as_pd(pt), I915_PDES, lvl);
227                         start += (u64)I915_PDES << gen8_pd_shift(lvl);
228                         continue;
229                 }
230
231                 if (lvl) {
232                         start = __gen8_ppgtt_clear(vm, as_pd(pt),
233                                                    start, end, lvl);
234                 } else {
235                         unsigned int count;
236                         u64 *vaddr;
237
238                         count = gen8_pt_count(start, end);
239                         DBG("%s(%p):{ lvl:%d, start:%llx, end:%llx, idx:%d, len:%d, used:%d } removing pte\n",
240                             __func__, vm, lvl, start, end,
241                             gen8_pd_index(start, 0), count,
242                             atomic_read(&pt->used));
243                         GEM_BUG_ON(!count || count >= atomic_read(&pt->used));
244
245                         vaddr = px_vaddr(pt);
246                         memset64(vaddr + gen8_pd_index(start, 0),
247                                  vm->scratch[0]->encode,
248                                  count);
249
250                         atomic_sub(count, &pt->used);
251                         start += count;
252                 }
253
254                 if (release_pd_entry(pd, idx, pt, scratch))
255                         free_px(vm, pt, lvl);
256         } while (idx++, --len);
257
258         return start;
259 }
260
261 static void gen8_ppgtt_clear(struct i915_address_space *vm,
262                              u64 start, u64 length)
263 {
264         GEM_BUG_ON(!IS_ALIGNED(start, BIT_ULL(GEN8_PTE_SHIFT)));
265         GEM_BUG_ON(!IS_ALIGNED(length, BIT_ULL(GEN8_PTE_SHIFT)));
266         GEM_BUG_ON(range_overflows(start, length, vm->total));
267
268         start >>= GEN8_PTE_SHIFT;
269         length >>= GEN8_PTE_SHIFT;
270         GEM_BUG_ON(length == 0);
271
272         __gen8_ppgtt_clear(vm, i915_vm_to_ppgtt(vm)->pd,
273                            start, start + length, vm->top);
274 }
275
276 static void __gen8_ppgtt_alloc(struct i915_address_space * const vm,
277                                struct i915_vm_pt_stash *stash,
278                                struct i915_page_directory * const pd,
279                                u64 * const start, const u64 end, int lvl)
280 {
281         unsigned int idx, len;
282
283         GEM_BUG_ON(end > vm->total >> GEN8_PTE_SHIFT);
284
285         len = gen8_pd_range(*start, end, lvl--, &idx);
286         DBG("%s(%p):{ lvl:%d, start:%llx, end:%llx, idx:%d, len:%d, used:%d }\n",
287             __func__, vm, lvl + 1, *start, end,
288             idx, len, atomic_read(px_used(pd)));
289         GEM_BUG_ON(!len || (idx + len - 1) >> gen8_pd_shift(1));
290
291         spin_lock(&pd->lock);
292         GEM_BUG_ON(!atomic_read(px_used(pd))); /* Must be pinned! */
293         do {
294                 struct i915_page_table *pt = pd->entry[idx];
295
296                 if (!pt) {
297                         spin_unlock(&pd->lock);
298
299                         DBG("%s(%p):{ lvl:%d, idx:%d } allocating new tree\n",
300                             __func__, vm, lvl + 1, idx);
301
302                         pt = stash->pt[!!lvl];
303                         __i915_gem_object_pin_pages(pt->base);
304                         i915_gem_object_make_unshrinkable(pt->base);
305
306                         fill_px(pt, vm->scratch[lvl]->encode);
307
308                         spin_lock(&pd->lock);
309                         if (likely(!pd->entry[idx])) {
310                                 stash->pt[!!lvl] = pt->stash;
311                                 atomic_set(&pt->used, 0);
312                                 set_pd_entry(pd, idx, pt);
313                         } else {
314                                 pt = pd->entry[idx];
315                         }
316                 }
317
318                 if (lvl) {
319                         atomic_inc(&pt->used);
320                         spin_unlock(&pd->lock);
321
322                         __gen8_ppgtt_alloc(vm, stash,
323                                            as_pd(pt), start, end, lvl);
324
325                         spin_lock(&pd->lock);
326                         atomic_dec(&pt->used);
327                         GEM_BUG_ON(!atomic_read(&pt->used));
328                 } else {
329                         unsigned int count = gen8_pt_count(*start, end);
330
331                         DBG("%s(%p):{ lvl:%d, start:%llx, end:%llx, idx:%d, len:%d, used:%d } inserting pte\n",
332                             __func__, vm, lvl, *start, end,
333                             gen8_pd_index(*start, 0), count,
334                             atomic_read(&pt->used));
335
336                         atomic_add(count, &pt->used);
337                         /* All other pdes may be simultaneously removed */
338                         GEM_BUG_ON(atomic_read(&pt->used) > NALLOC * I915_PDES);
339                         *start += count;
340                 }
341         } while (idx++, --len);
342         spin_unlock(&pd->lock);
343 }
344
345 static void gen8_ppgtt_alloc(struct i915_address_space *vm,
346                              struct i915_vm_pt_stash *stash,
347                              u64 start, u64 length)
348 {
349         GEM_BUG_ON(!IS_ALIGNED(start, BIT_ULL(GEN8_PTE_SHIFT)));
350         GEM_BUG_ON(!IS_ALIGNED(length, BIT_ULL(GEN8_PTE_SHIFT)));
351         GEM_BUG_ON(range_overflows(start, length, vm->total));
352
353         start >>= GEN8_PTE_SHIFT;
354         length >>= GEN8_PTE_SHIFT;
355         GEM_BUG_ON(length == 0);
356
357         __gen8_ppgtt_alloc(vm, stash, i915_vm_to_ppgtt(vm)->pd,
358                            &start, start + length, vm->top);
359 }
360
361 static void __gen8_ppgtt_foreach(struct i915_address_space *vm,
362                                  struct i915_page_directory *pd,
363                                  u64 *start, u64 end, int lvl,
364                                  void (*fn)(struct i915_address_space *vm,
365                                             struct i915_page_table *pt,
366                                             void *data),
367                                  void *data)
368 {
369         unsigned int idx, len;
370
371         len = gen8_pd_range(*start, end, lvl--, &idx);
372
373         spin_lock(&pd->lock);
374         do {
375                 struct i915_page_table *pt = pd->entry[idx];
376
377                 atomic_inc(&pt->used);
378                 spin_unlock(&pd->lock);
379
380                 if (lvl) {
381                         __gen8_ppgtt_foreach(vm, as_pd(pt), start, end, lvl,
382                                              fn, data);
383                 } else {
384                         fn(vm, pt, data);
385                         *start += gen8_pt_count(*start, end);
386                 }
387
388                 spin_lock(&pd->lock);
389                 atomic_dec(&pt->used);
390         } while (idx++, --len);
391         spin_unlock(&pd->lock);
392 }
393
394 static void gen8_ppgtt_foreach(struct i915_address_space *vm,
395                                u64 start, u64 length,
396                                void (*fn)(struct i915_address_space *vm,
397                                           struct i915_page_table *pt,
398                                           void *data),
399                                void *data)
400 {
401         start >>= GEN8_PTE_SHIFT;
402         length >>= GEN8_PTE_SHIFT;
403
404         __gen8_ppgtt_foreach(vm, i915_vm_to_ppgtt(vm)->pd,
405                              &start, start + length, vm->top,
406                              fn, data);
407 }
408
409 static __always_inline u64
410 gen8_ppgtt_insert_pte(struct i915_ppgtt *ppgtt,
411                       struct i915_page_directory *pdp,
412                       struct sgt_dma *iter,
413                       u64 idx,
414                       enum i915_cache_level cache_level,
415                       u32 flags)
416 {
417         struct i915_page_directory *pd;
418         const gen8_pte_t pte_encode = gen8_pte_encode(0, cache_level, flags);
419         gen8_pte_t *vaddr;
420
421         pd = i915_pd_entry(pdp, gen8_pd_index(idx, 2));
422         vaddr = px_vaddr(i915_pt_entry(pd, gen8_pd_index(idx, 1)));
423         do {
424                 GEM_BUG_ON(sg_dma_len(iter->sg) < I915_GTT_PAGE_SIZE);
425                 vaddr[gen8_pd_index(idx, 0)] = pte_encode | iter->dma;
426
427                 iter->dma += I915_GTT_PAGE_SIZE;
428                 if (iter->dma >= iter->max) {
429                         iter->sg = __sg_next(iter->sg);
430                         if (!iter->sg || sg_dma_len(iter->sg) == 0) {
431                                 idx = 0;
432                                 break;
433                         }
434
435                         iter->dma = sg_dma_address(iter->sg);
436                         iter->max = iter->dma + sg_dma_len(iter->sg);
437                 }
438
439                 if (gen8_pd_index(++idx, 0) == 0) {
440                         if (gen8_pd_index(idx, 1) == 0) {
441                                 /* Limited by sg length for 3lvl */
442                                 if (gen8_pd_index(idx, 2) == 0)
443                                         break;
444
445                                 pd = pdp->entry[gen8_pd_index(idx, 2)];
446                         }
447
448                         clflush_cache_range(vaddr, PAGE_SIZE);
449                         vaddr = px_vaddr(i915_pt_entry(pd, gen8_pd_index(idx, 1)));
450                 }
451         } while (1);
452         clflush_cache_range(vaddr, PAGE_SIZE);
453
454         return idx;
455 }
456
457 static void gen8_ppgtt_insert_huge(struct i915_vma *vma,
458                                    struct sgt_dma *iter,
459                                    enum i915_cache_level cache_level,
460                                    u32 flags)
461 {
462         const gen8_pte_t pte_encode = gen8_pte_encode(0, cache_level, flags);
463         unsigned int rem = sg_dma_len(iter->sg);
464         u64 start = vma->node.start;
465
466         GEM_BUG_ON(!i915_vm_is_4lvl(vma->vm));
467
468         do {
469                 struct i915_page_directory * const pdp =
470                         gen8_pdp_for_page_address(vma->vm, start);
471                 struct i915_page_directory * const pd =
472                         i915_pd_entry(pdp, __gen8_pte_index(start, 2));
473                 gen8_pte_t encode = pte_encode;
474                 unsigned int maybe_64K = -1;
475                 unsigned int page_size;
476                 gen8_pte_t *vaddr;
477                 u16 index;
478
479                 if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_2M &&
480                     IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_2M) &&
481                     rem >= I915_GTT_PAGE_SIZE_2M &&
482                     !__gen8_pte_index(start, 0)) {
483                         index = __gen8_pte_index(start, 1);
484                         encode |= GEN8_PDE_PS_2M;
485                         page_size = I915_GTT_PAGE_SIZE_2M;
486
487                         vaddr = px_vaddr(pd);
488                 } else {
489                         struct i915_page_table *pt =
490                                 i915_pt_entry(pd, __gen8_pte_index(start, 1));
491
492                         index = __gen8_pte_index(start, 0);
493                         page_size = I915_GTT_PAGE_SIZE;
494
495                         if (!index &&
496                             vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K &&
497                             IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_64K) &&
498                             (IS_ALIGNED(rem, I915_GTT_PAGE_SIZE_64K) ||
499                              rem >= (I915_PDES - index) * I915_GTT_PAGE_SIZE))
500                                 maybe_64K = __gen8_pte_index(start, 1);
501
502                         vaddr = px_vaddr(pt);
503                 }
504
505                 do {
506                         GEM_BUG_ON(sg_dma_len(iter->sg) < page_size);
507                         vaddr[index++] = encode | iter->dma;
508
509                         start += page_size;
510                         iter->dma += page_size;
511                         rem -= page_size;
512                         if (iter->dma >= iter->max) {
513                                 iter->sg = __sg_next(iter->sg);
514                                 if (!iter->sg)
515                                         break;
516
517                                 rem = sg_dma_len(iter->sg);
518                                 if (!rem)
519                                         break;
520
521                                 iter->dma = sg_dma_address(iter->sg);
522                                 iter->max = iter->dma + rem;
523
524                                 if (maybe_64K != -1 && index < I915_PDES &&
525                                     !(IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_64K) &&
526                                       (IS_ALIGNED(rem, I915_GTT_PAGE_SIZE_64K) ||
527                                        rem >= (I915_PDES - index) * I915_GTT_PAGE_SIZE)))
528                                         maybe_64K = -1;
529
530                                 if (unlikely(!IS_ALIGNED(iter->dma, page_size)))
531                                         break;
532                         }
533                 } while (rem >= page_size && index < I915_PDES);
534
535                 clflush_cache_range(vaddr, PAGE_SIZE);
536
537                 /*
538                  * Is it safe to mark the 2M block as 64K? -- Either we have
539                  * filled whole page-table with 64K entries, or filled part of
540                  * it and have reached the end of the sg table and we have
541                  * enough padding.
542                  */
543                 if (maybe_64K != -1 &&
544                     (index == I915_PDES ||
545                      (i915_vm_has_scratch_64K(vma->vm) &&
546                       !iter->sg && IS_ALIGNED(vma->node.start +
547                                               vma->node.size,
548                                               I915_GTT_PAGE_SIZE_2M)))) {
549                         vaddr = px_vaddr(pd);
550                         vaddr[maybe_64K] |= GEN8_PDE_IPS_64K;
551                         page_size = I915_GTT_PAGE_SIZE_64K;
552
553                         /*
554                          * We write all 4K page entries, even when using 64K
555                          * pages. In order to verify that the HW isn't cheating
556                          * by using the 4K PTE instead of the 64K PTE, we want
557                          * to remove all the surplus entries. If the HW skipped
558                          * the 64K PTE, it will read/write into the scratch page
559                          * instead - which we detect as missing results during
560                          * selftests.
561                          */
562                         if (I915_SELFTEST_ONLY(vma->vm->scrub_64K)) {
563                                 u16 i;
564
565                                 encode = vma->vm->scratch[0]->encode;
566                                 vaddr = px_vaddr(i915_pt_entry(pd, maybe_64K));
567
568                                 for (i = 1; i < index; i += 16)
569                                         memset64(vaddr + i, encode, 15);
570
571                         }
572                 }
573
574                 vma->page_sizes.gtt |= page_size;
575         } while (iter->sg && sg_dma_len(iter->sg));
576 }
577
578 static void gen8_ppgtt_insert(struct i915_address_space *vm,
579                               struct i915_vma *vma,
580                               enum i915_cache_level cache_level,
581                               u32 flags)
582 {
583         struct i915_ppgtt * const ppgtt = i915_vm_to_ppgtt(vm);
584         struct sgt_dma iter = sgt_dma(vma);
585
586         if (vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
587                 gen8_ppgtt_insert_huge(vma, &iter, cache_level, flags);
588         } else  {
589                 u64 idx = vma->node.start >> GEN8_PTE_SHIFT;
590
591                 do {
592                         struct i915_page_directory * const pdp =
593                                 gen8_pdp_for_page_index(vm, idx);
594
595                         idx = gen8_ppgtt_insert_pte(ppgtt, pdp, &iter, idx,
596                                                     cache_level, flags);
597                 } while (idx);
598
599                 vma->page_sizes.gtt = I915_GTT_PAGE_SIZE;
600         }
601 }
602
603 static void gen8_ppgtt_insert_entry(struct i915_address_space *vm,
604                                     dma_addr_t addr,
605                                     u64 offset,
606                                     enum i915_cache_level level,
607                                     u32 flags)
608 {
609         u64 idx = offset >> GEN8_PTE_SHIFT;
610         struct i915_page_directory * const pdp =
611                 gen8_pdp_for_page_index(vm, idx);
612         struct i915_page_directory *pd =
613                 i915_pd_entry(pdp, gen8_pd_index(idx, 2));
614         gen8_pte_t *vaddr;
615
616         vaddr = px_vaddr(i915_pt_entry(pd, gen8_pd_index(idx, 1)));
617         vaddr[gen8_pd_index(idx, 0)] = gen8_pte_encode(addr, level, flags);
618         clflush_cache_range(&vaddr[gen8_pd_index(idx, 0)], sizeof(*vaddr));
619 }
620
621 static int gen8_init_scratch(struct i915_address_space *vm)
622 {
623         u32 pte_flags;
624         int ret;
625         int i;
626
627         /*
628          * If everybody agrees to not to write into the scratch page,
629          * we can reuse it for all vm, keeping contexts and processes separate.
630          */
631         if (vm->has_read_only && vm->gt->vm && !i915_is_ggtt(vm->gt->vm)) {
632                 struct i915_address_space *clone = vm->gt->vm;
633
634                 GEM_BUG_ON(!clone->has_read_only);
635
636                 vm->scratch_order = clone->scratch_order;
637                 for (i = 0; i <= vm->top; i++)
638                         vm->scratch[i] = i915_gem_object_get(clone->scratch[i]);
639
640                 return 0;
641         }
642
643         ret = setup_scratch_page(vm);
644         if (ret)
645                 return ret;
646
647         pte_flags = vm->has_read_only;
648         if (i915_gem_object_is_lmem(vm->scratch[0]))
649                 pte_flags |= PTE_LM;
650
651         vm->scratch[0]->encode =
652                 gen8_pte_encode(px_dma(vm->scratch[0]),
653                                 I915_CACHE_LLC, pte_flags);
654
655         for (i = 1; i <= vm->top; i++) {
656                 struct drm_i915_gem_object *obj;
657
658                 obj = vm->alloc_pt_dma(vm, I915_GTT_PAGE_SIZE_4K);
659                 if (IS_ERR(obj))
660                         goto free_scratch;
661
662                 ret = map_pt_dma(vm, obj);
663                 if (ret) {
664                         i915_gem_object_put(obj);
665                         goto free_scratch;
666                 }
667
668                 fill_px(obj, vm->scratch[i - 1]->encode);
669                 obj->encode = gen8_pde_encode(px_dma(obj), I915_CACHE_LLC);
670
671                 vm->scratch[i] = obj;
672         }
673
674         return 0;
675
676 free_scratch:
677         while (i--)
678                 i915_gem_object_put(vm->scratch[i]);
679         return -ENOMEM;
680 }
681
682 static int gen8_preallocate_top_level_pdp(struct i915_ppgtt *ppgtt)
683 {
684         struct i915_address_space *vm = &ppgtt->vm;
685         struct i915_page_directory *pd = ppgtt->pd;
686         unsigned int idx;
687
688         GEM_BUG_ON(vm->top != 2);
689         GEM_BUG_ON(gen8_pd_top_count(vm) != GEN8_3LVL_PDPES);
690
691         for (idx = 0; idx < GEN8_3LVL_PDPES; idx++) {
692                 struct i915_page_directory *pde;
693                 int err;
694
695                 pde = alloc_pd(vm);
696                 if (IS_ERR(pde))
697                         return PTR_ERR(pde);
698
699                 err = map_pt_dma(vm, pde->pt.base);
700                 if (err) {
701                         free_pd(vm, pde);
702                         return err;
703                 }
704
705                 fill_px(pde, vm->scratch[1]->encode);
706                 set_pd_entry(pd, idx, pde);
707                 atomic_inc(px_used(pde)); /* keep pinned */
708         }
709         wmb();
710
711         return 0;
712 }
713
714 static struct i915_page_directory *
715 gen8_alloc_top_pd(struct i915_address_space *vm)
716 {
717         const unsigned int count = gen8_pd_top_count(vm);
718         struct i915_page_directory *pd;
719         int err;
720
721         GEM_BUG_ON(count > I915_PDES);
722
723         pd = __alloc_pd(count);
724         if (unlikely(!pd))
725                 return ERR_PTR(-ENOMEM);
726
727         pd->pt.base = vm->alloc_pt_dma(vm, I915_GTT_PAGE_SIZE_4K);
728         if (IS_ERR(pd->pt.base)) {
729                 err = PTR_ERR(pd->pt.base);
730                 pd->pt.base = NULL;
731                 goto err_pd;
732         }
733
734         err = map_pt_dma(vm, pd->pt.base);
735         if (err)
736                 goto err_pd;
737
738         fill_page_dma(px_base(pd), vm->scratch[vm->top]->encode, count);
739         atomic_inc(px_used(pd)); /* mark as pinned */
740         return pd;
741
742 err_pd:
743         free_pd(vm, pd);
744         return ERR_PTR(err);
745 }
746
747 /*
748  * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
749  * with a net effect resembling a 2-level page table in normal x86 terms. Each
750  * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
751  * space.
752  *
753  */
754 struct i915_ppgtt *gen8_ppgtt_create(struct intel_gt *gt)
755 {
756         struct i915_ppgtt *ppgtt;
757         int err;
758
759         ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
760         if (!ppgtt)
761                 return ERR_PTR(-ENOMEM);
762
763         ppgtt_init(ppgtt, gt);
764         ppgtt->vm.top = i915_vm_is_4lvl(&ppgtt->vm) ? 3 : 2;
765         ppgtt->vm.pd_shift = ilog2(SZ_4K * SZ_4K / sizeof(gen8_pte_t));
766
767         /*
768          * From bdw, there is hw support for read-only pages in the PPGTT.
769          *
770          * Gen11 has HSDES#:1807136187 unresolved. Disable ro support
771          * for now.
772          *
773          * Gen12 has inherited the same read-only fault issue from gen11.
774          */
775         ppgtt->vm.has_read_only = !IS_GRAPHICS_VER(gt->i915, 11, 12);
776
777         if (HAS_LMEM(gt->i915))
778                 ppgtt->vm.alloc_pt_dma = alloc_pt_lmem;
779         else
780                 ppgtt->vm.alloc_pt_dma = alloc_pt_dma;
781
782         err = gen8_init_scratch(&ppgtt->vm);
783         if (err)
784                 goto err_free;
785
786         ppgtt->pd = gen8_alloc_top_pd(&ppgtt->vm);
787         if (IS_ERR(ppgtt->pd)) {
788                 err = PTR_ERR(ppgtt->pd);
789                 goto err_free_scratch;
790         }
791
792         if (!i915_vm_is_4lvl(&ppgtt->vm)) {
793                 err = gen8_preallocate_top_level_pdp(ppgtt);
794                 if (err)
795                         goto err_free_pd;
796         }
797
798         ppgtt->vm.bind_async_flags = I915_VMA_LOCAL_BIND;
799         ppgtt->vm.insert_entries = gen8_ppgtt_insert;
800         ppgtt->vm.insert_page = gen8_ppgtt_insert_entry;
801         ppgtt->vm.allocate_va_range = gen8_ppgtt_alloc;
802         ppgtt->vm.clear_range = gen8_ppgtt_clear;
803         ppgtt->vm.foreach = gen8_ppgtt_foreach;
804
805         ppgtt->vm.pte_encode = gen8_pte_encode;
806
807         if (intel_vgpu_active(gt->i915))
808                 gen8_ppgtt_notify_vgt(ppgtt, true);
809
810         ppgtt->vm.cleanup = gen8_ppgtt_cleanup;
811
812         return ppgtt;
813
814 err_free_pd:
815         __gen8_ppgtt_cleanup(&ppgtt->vm, ppgtt->pd,
816                              gen8_pd_top_count(&ppgtt->vm), ppgtt->vm.top);
817 err_free_scratch:
818         free_scratch(&ppgtt->vm);
819 err_free:
820         kfree(ppgtt);
821         return ERR_PTR(err);
822 }