gt/gen6_ppgtt.c

   1 // SPDX-License-Identifier: MIT
   2 /*
   3  * Copyright © 2020 Intel Corporation
   4  */
   5
   6 #include <linux/log2.h>
   7
   8 #include "gen6_ppgtt.h"
   9 #include "i915_scatterlist.h"
  10 #include "i915_trace.h"
  11 #include "i915_vgpu.h"
  12 #include "intel_gt.h"
  13
  14 /* Write pde (index) from the page directory @pd to the page table @pt */
  15 static inline void gen6_write_pde(const struct gen6_ppgtt *ppgtt,
  16                                   const unsigned int pde,
  17                                   const struct i915_page_table *pt)
  18 {
  19         dma_addr_t addr = pt ? px_dma(pt) : px_dma(ppgtt->base.vm.scratch[1]);
  20
  21         /* Caller needs to make sure the write completes if necessary */
  22         iowrite32(GEN6_PDE_ADDR_ENCODE(addr) | GEN6_PDE_VALID,
  23                   ppgtt->pd_addr + pde);
  24 }
  25
  26 void gen7_ppgtt_enable(struct intel_gt *gt)
  27 {
  28         struct drm_i915_private *i915 = gt->i915;
  29         struct intel_uncore *uncore = gt->uncore;
  30         struct intel_engine_cs *engine;
  31         enum intel_engine_id id;
  32         u32 ecochk;
  33
  34         intel_uncore_rmw(uncore, GAC_ECO_BITS, 0, ECOBITS_PPGTT_CACHE64B);
  35
  36         ecochk = intel_uncore_read(uncore, GAM_ECOCHK);
  37         if (IS_HASWELL(i915)) {
  38                 ecochk |= ECOCHK_PPGTT_WB_HSW;
  39         } else {
  40                 ecochk |= ECOCHK_PPGTT_LLC_IVB;
  41                 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
  42         }
  43         intel_uncore_write(uncore, GAM_ECOCHK, ecochk);
  44
  45         for_each_engine(engine, gt, id) {
  46                 /* GFX_MODE is per-ring on gen7+ */
  47                 ENGINE_WRITE(engine,
  48                              RING_MODE_GEN7,
  49                              _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
  50         }
  51 }
  52
  53 void gen6_ppgtt_enable(struct intel_gt *gt)
  54 {
  55         struct intel_uncore *uncore = gt->uncore;
  56
  57         intel_uncore_rmw(uncore,
  58                          GAC_ECO_BITS,
  59                          0,
  60                          ECOBITS_SNB_BIT | ECOBITS_PPGTT_CACHE64B);
  61
  62         intel_uncore_rmw(uncore,
  63                          GAB_CTL,
  64                          0,
  65                          GAB_CTL_CONT_AFTER_PAGEFAULT);
  66
  67         intel_uncore_rmw(uncore,
  68                          GAM_ECOCHK,
  69                          0,
  70                          ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
  71
  72         if (HAS_PPGTT(uncore->i915)) /* may be disabled for VT-d */
  73                 intel_uncore_write(uncore,
  74                                    GFX_MODE,
  75                                    _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
  76 }
  77
  78 /* PPGTT support for Sandybdrige/Gen6 and later */
  79 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
  80                                    u64 start, u64 length)
  81 {
  82         struct gen6_ppgtt * const ppgtt = to_gen6_ppgtt(i915_vm_to_ppgtt(vm));
  83         const unsigned int first_entry = start / I915_GTT_PAGE_SIZE;
  84         const gen6_pte_t scratch_pte = vm->scratch[0]->encode;
  85         unsigned int pde = first_entry / GEN6_PTES;
  86         unsigned int pte = first_entry % GEN6_PTES;
  87         unsigned int num_entries = length / I915_GTT_PAGE_SIZE;
  88
  89         while (num_entries) {
  90                 struct i915_page_table * const pt =
  91                         i915_pt_entry(ppgtt->base.pd, pde++);
  92                 const unsigned int count = min(num_entries, GEN6_PTES - pte);
  93                 gen6_pte_t *vaddr;
  94
  95                 num_entries -= count;
  96
  97                 GEM_BUG_ON(count > atomic_read(&pt->used));
  98                 if (!atomic_sub_return(count, &pt->used))
  99                         ppgtt->scan_for_unused_pt = true;
 100
 101                 /*
 102                  * Note that the hw doesn't support removing PDE on the fly
 103                  * (they are cached inside the context with no means to
 104                  * invalidate the cache), so we can only reset the PTE
 105                  * entries back to scratch.
 106                  */
 107
 108                 vaddr = kmap_atomic_px(pt);
 109                 memset32(vaddr + pte, scratch_pte, count);
 110                 kunmap_atomic(vaddr);
 111
 112                 pte = 0;
 113         }
 114 }
 115
 116 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
 117                                       struct i915_vma *vma,
 118                                       enum i915_cache_level cache_level,
 119                                       u32 flags)
 120 {
 121         struct i915_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
 122         struct i915_page_directory * const pd = ppgtt->pd;
 123         unsigned int first_entry = vma->node.start / I915_GTT_PAGE_SIZE;
 124         unsigned int act_pt = first_entry / GEN6_PTES;
 125         unsigned int act_pte = first_entry % GEN6_PTES;
 126         const u32 pte_encode = vm->pte_encode(0, cache_level, flags);
 127         struct sgt_dma iter = sgt_dma(vma);
 128         gen6_pte_t *vaddr;
 129
 130         GEM_BUG_ON(!pd->entry[act_pt]);
 131
 132         vaddr = kmap_atomic_px(i915_pt_entry(pd, act_pt));
 133         do {
 134                 GEM_BUG_ON(iter.sg->length < I915_GTT_PAGE_SIZE);
 135                 vaddr[act_pte] = pte_encode | GEN6_PTE_ADDR_ENCODE(iter.dma);
 136
 137                 iter.dma += I915_GTT_PAGE_SIZE;
 138                 if (iter.dma == iter.max) {
 139                         iter.sg = __sg_next(iter.sg);
 140                         if (!iter.sg)
 141                                 break;
 142
 143                         iter.dma = sg_dma_address(iter.sg);
 144                         iter.max = iter.dma + iter.sg->length;
 145                 }
 146
 147                 if (++act_pte == GEN6_PTES) {
 148                         kunmap_atomic(vaddr);
 149                         vaddr = kmap_atomic_px(i915_pt_entry(pd, ++act_pt));
 150                         act_pte = 0;
 151                 }
 152         } while (1);
 153         kunmap_atomic(vaddr);
 154
 155         vma->page_sizes.gtt = I915_GTT_PAGE_SIZE;
 156 }
 157
 158 static void gen6_flush_pd(struct gen6_ppgtt *ppgtt, u64 start, u64 end)
 159 {
 160         struct i915_page_directory * const pd = ppgtt->base.pd;
 161         struct i915_page_table *pt;
 162         unsigned int pde;
 163
 164         start = round_down(start, SZ_64K);
 165         end = round_up(end, SZ_64K) - start;
 166
 167         mutex_lock(&ppgtt->flush);
 168
 169         gen6_for_each_pde(pt, pd, start, end, pde)
 170                 gen6_write_pde(ppgtt, pde, pt);
 171
 172         mb();
 173         ioread32(ppgtt->pd_addr + pde - 1);
 174         gen6_ggtt_invalidate(ppgtt->base.vm.gt->ggtt);
 175         mb();
 176
 177         mutex_unlock(&ppgtt->flush);
 178 }
 179
 180 static void gen6_alloc_va_range(struct i915_address_space *vm,
 181                                 struct i915_vm_pt_stash *stash,
 182                                 u64 start, u64 length)
 183 {
 184         struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(i915_vm_to_ppgtt(vm));
 185         struct i915_page_directory * const pd = ppgtt->base.pd;
 186         struct i915_page_table *pt;
 187         bool flush = false;
 188         u64 from = start;
 189         unsigned int pde;
 190
 191         spin_lock(&pd->lock);
 192         gen6_for_each_pde(pt, pd, start, length, pde) {
 193                 const unsigned int count = gen6_pte_count(start, length);
 194
 195                 if (!pt) {
 196                         spin_unlock(&pd->lock);
 197
 198                         pt = stash->pt[0];
 199                         __i915_gem_object_pin_pages(pt->base);
 200                         i915_gem_object_make_unshrinkable(pt->base);
 201
 202                         fill32_px(pt, vm->scratch[0]->encode);
 203
 204                         spin_lock(&pd->lock);
 205                         if (!pd->entry[pde]) {
 206                                 stash->pt[0] = pt->stash;
 207                                 atomic_set(&pt->used, 0);
 208                                 pd->entry[pde] = pt;
 209                         } else {
 210                                 pt = pd->entry[pde];
 211                         }
 212
 213                         flush = true;
 214                 }
 215
 216                 atomic_add(count, &pt->used);
 217         }
 218         spin_unlock(&pd->lock);
 219
 220         if (flush && i915_vma_is_bound(ppgtt->vma, I915_VMA_GLOBAL_BIND)) {
 221                 intel_wakeref_t wakeref;
 222
 223                 with_intel_runtime_pm(&vm->i915->runtime_pm, wakeref)
 224                         gen6_flush_pd(ppgtt, from, start);
 225         }
 226 }
 227
 228 static int gen6_ppgtt_init_scratch(struct gen6_ppgtt *ppgtt)
 229 {
 230         struct i915_address_space * const vm = &ppgtt->base.vm;
 231         int ret;
 232
 233         ret = setup_scratch_page(vm);
 234         if (ret)
 235                 return ret;
 236
 237         vm->scratch[0]->encode =
 238                 vm->pte_encode(px_dma(vm->scratch[0]),
 239                                I915_CACHE_NONE, PTE_READ_ONLY);
 240
 241         vm->scratch[1] = vm->alloc_pt_dma(vm, I915_GTT_PAGE_SIZE_4K);
 242         if (IS_ERR(vm->scratch[1])) {
 243                 ret = PTR_ERR(vm->scratch[1]);
 244                 goto err_scratch0;
 245         }
 246
 247         ret = pin_pt_dma(vm, vm->scratch[1]);
 248         if (ret)
 249                 goto err_scratch1;
 250
 251         fill32_px(vm->scratch[1], vm->scratch[0]->encode);
 252
 253         return 0;
 254
 255 err_scratch1:
 256         i915_gem_object_put(vm->scratch[1]);
 257 err_scratch0:
 258         i915_gem_object_put(vm->scratch[0]);
 259         return ret;
 260 }
 261
 262 static void gen6_ppgtt_free_pd(struct gen6_ppgtt *ppgtt)
 263 {
 264         struct i915_page_directory * const pd = ppgtt->base.pd;
 265         struct i915_page_table *pt;
 266         u32 pde;
 267
 268         gen6_for_all_pdes(pt, pd, pde)
 269                 if (pt)
 270                         free_pt(&ppgtt->base.vm, pt);
 271 }
 272
 273 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
 274 {
 275         struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(i915_vm_to_ppgtt(vm));
 276
 277         __i915_vma_put(ppgtt->vma);
 278
 279         gen6_ppgtt_free_pd(ppgtt);
 280         free_scratch(vm);
 281
 282         mutex_destroy(&ppgtt->flush);
 283         mutex_destroy(&ppgtt->pin_mutex);
 284
 285         free_pd(&ppgtt->base.vm, ppgtt->base.pd);
 286 }
 287
 288 static int pd_vma_set_pages(struct i915_vma *vma)
 289 {
 290         vma->pages = ERR_PTR(-ENODEV);
 291         return 0;
 292 }
 293
 294 static void pd_vma_clear_pages(struct i915_vma *vma)
 295 {
 296         GEM_BUG_ON(!vma->pages);
 297
 298         vma->pages = NULL;
 299 }
 300
 301 static void pd_vma_bind(struct i915_address_space *vm,
 302                         struct i915_vm_pt_stash *stash,
 303                         struct i915_vma *vma,
 304                         enum i915_cache_level cache_level,
 305                         u32 unused)
 306 {
 307         struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
 308         struct gen6_ppgtt *ppgtt = vma->private;
 309         u32 ggtt_offset = i915_ggtt_offset(vma) / I915_GTT_PAGE_SIZE;
 310
 311         ppgtt->pp_dir = ggtt_offset * sizeof(gen6_pte_t) << 10;
 312         ppgtt->pd_addr = (gen6_pte_t __iomem *)ggtt->gsm + ggtt_offset;
 313
 314         gen6_flush_pd(ppgtt, 0, ppgtt->base.vm.total);
 315 }
 316
 317 static void pd_vma_unbind(struct i915_address_space *vm, struct i915_vma *vma)
 318 {
 319         struct gen6_ppgtt *ppgtt = vma->private;
 320         struct i915_page_directory * const pd = ppgtt->base.pd;
 321         struct i915_page_table *pt;
 322         unsigned int pde;
 323
 324         if (!ppgtt->scan_for_unused_pt)
 325                 return;
 326
 327         /* Free all no longer used page tables */
 328         gen6_for_all_pdes(pt, ppgtt->base.pd, pde) {
 329                 if (!pt || atomic_read(&pt->used))
 330                         continue;
 331
 332                 free_pt(&ppgtt->base.vm, pt);
 333                 pd->entry[pde] = NULL;
 334         }
 335
 336         ppgtt->scan_for_unused_pt = false;
 337 }
 338
 339 static const struct i915_vma_ops pd_vma_ops = {
 340         .set_pages = pd_vma_set_pages,
 341         .clear_pages = pd_vma_clear_pages,
 342         .bind_vma = pd_vma_bind,
 343         .unbind_vma = pd_vma_unbind,
 344 };
 345
 346 static struct i915_vma *pd_vma_create(struct gen6_ppgtt *ppgtt, int size)
 347 {
 348         struct i915_ggtt *ggtt = ppgtt->base.vm.gt->ggtt;
 349         struct i915_vma *vma;
 350
 351         GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
 352         GEM_BUG_ON(size > ggtt->vm.total);
 353
 354         vma = i915_vma_alloc();
 355         if (!vma)
 356                 return ERR_PTR(-ENOMEM);
 357
 358         i915_active_init(&vma->active, NULL, NULL);
 359
 360         kref_init(&vma->ref);
 361         mutex_init(&vma->pages_mutex);
 362         vma->vm = i915_vm_get(&ggtt->vm);
 363         vma->ops = &pd_vma_ops;
 364         vma->private = ppgtt;
 365
 366         vma->size = size;
 367         vma->fence_size = size;
 368         atomic_set(&vma->flags, I915_VMA_GGTT);
 369         vma->ggtt_view.type = I915_GGTT_VIEW_ROTATED; /* prevent fencing */
 370
 371         INIT_LIST_HEAD(&vma->obj_link);
 372         INIT_LIST_HEAD(&vma->closed_link);
 373
 374         return vma;
 375 }
 376
 377 int gen6_ppgtt_pin(struct i915_ppgtt *base, struct i915_gem_ww_ctx *ww)
 378 {
 379         struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(base);
 380         int err;
 381
 382         GEM_BUG_ON(!atomic_read(&ppgtt->base.vm.open));
 383
 384         /*
 385          * Workaround the limited maximum vma->pin_count and the aliasing_ppgtt
 386          * which will be pinned into every active context.
 387          * (When vma->pin_count becomes atomic, I expect we will naturally
 388          * need a larger, unpacked, type and kill this redundancy.)
 389          */
 390         if (atomic_add_unless(&ppgtt->pin_count, 1, 0))
 391                 return 0;
 392
 393         if (mutex_lock_interruptible(&ppgtt->pin_mutex))
 394                 return -EINTR;
 395
 396         /*
 397          * PPGTT PDEs reside in the GGTT and consists of 512 entries. The
 398          * allocator works in address space sizes, so it's multiplied by page
 399          * size. We allocate at the top of the GTT to avoid fragmentation.
 400          */
 401         err = 0;
 402         if (!atomic_read(&ppgtt->pin_count))
 403                 err = i915_ggtt_pin(ppgtt->vma, ww, GEN6_PD_ALIGN, PIN_HIGH);
 404         if (!err)
 405                 atomic_inc(&ppgtt->pin_count);
 406         mutex_unlock(&ppgtt->pin_mutex);
 407
 408         return err;
 409 }
 410
 411 void gen6_ppgtt_unpin(struct i915_ppgtt *base)
 412 {
 413         struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(base);
 414
 415         GEM_BUG_ON(!atomic_read(&ppgtt->pin_count));
 416         if (atomic_dec_and_test(&ppgtt->pin_count))
 417                 i915_vma_unpin(ppgtt->vma);
 418 }
 419
 420 void gen6_ppgtt_unpin_all(struct i915_ppgtt *base)
 421 {
 422         struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(base);
 423
 424         if (!atomic_read(&ppgtt->pin_count))
 425                 return;
 426
 427         i915_vma_unpin(ppgtt->vma);
 428         atomic_set(&ppgtt->pin_count, 0);
 429 }
 430
 431 struct i915_ppgtt *gen6_ppgtt_create(struct intel_gt *gt)
 432 {
 433         struct i915_ggtt * const ggtt = gt->ggtt;
 434         struct gen6_ppgtt *ppgtt;
 435         int err;
 436
 437         ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
 438         if (!ppgtt)
 439                 return ERR_PTR(-ENOMEM);
 440
 441         mutex_init(&ppgtt->flush);
 442         mutex_init(&ppgtt->pin_mutex);
 443
 444         ppgtt_init(&ppgtt->base, gt);
 445         ppgtt->base.vm.pd_shift = ilog2(SZ_4K * SZ_4K / sizeof(gen6_pte_t));
 446         ppgtt->base.vm.top = 1;
 447
 448         ppgtt->base.vm.bind_async_flags = I915_VMA_LOCAL_BIND;
 449         ppgtt->base.vm.allocate_va_range = gen6_alloc_va_range;
 450         ppgtt->base.vm.clear_range = gen6_ppgtt_clear_range;
 451         ppgtt->base.vm.insert_entries = gen6_ppgtt_insert_entries;
 452         ppgtt->base.vm.cleanup = gen6_ppgtt_cleanup;
 453
 454         ppgtt->base.vm.alloc_pt_dma = alloc_pt_dma;
 455         ppgtt->base.vm.pte_encode = ggtt->vm.pte_encode;
 456
 457         ppgtt->base.pd = __alloc_pd(I915_PDES);
 458         if (!ppgtt->base.pd) {
 459                 err = -ENOMEM;
 460                 goto err_free;
 461         }
 462
 463         err = gen6_ppgtt_init_scratch(ppgtt);
 464         if (err)
 465                 goto err_pd;
 466
 467         ppgtt->vma = pd_vma_create(ppgtt, GEN6_PD_SIZE);
 468         if (IS_ERR(ppgtt->vma)) {
 469                 err = PTR_ERR(ppgtt->vma);
 470                 goto err_scratch;
 471         }
 472
 473         return &ppgtt->base;
 474
 475 err_scratch:
 476         free_scratch(&ppgtt->base.vm);
 477 err_pd:
 478         free_pd(&ppgtt->base.vm, ppgtt->base.pd);
 479 err_free:
 480         mutex_destroy(&ppgtt->pin_mutex);
 481         kfree(ppgtt);
 482         return ERR_PTR(err);
 483 }