| 1 | // SPDX-License-Identifier: MIT |
| 2 | /* |
| 3 | * Copyright © 2014-2016 Intel Corporation |
| 4 | */ |
| 5 | |
| 6 | #include <linux/vmalloc.h> |
| 7 | |
| 8 | #include <drm/drm_cache.h> |
| 9 | #include <drm/drm_panic.h> |
| 10 | #include <drm/drm_print.h> |
| 11 | |
| 12 | #include "display/intel_fb.h" |
| 13 | #include "display/intel_display_types.h" |
| 14 | #include "gt/intel_gt.h" |
| 15 | #include "gt/intel_tlb.h" |
| 16 | |
| 17 | #include "i915_drv.h" |
| 18 | #include "i915_gem_object.h" |
| 19 | #include "i915_scatterlist.h" |
| 20 | #include "i915_gem_lmem.h" |
| 21 | #include "i915_gem_mman.h" |
| 22 | |
| 23 | void __i915_gem_object_set_pages(struct drm_i915_gem_object *obj, |
| 24 | struct sg_table *pages) |
| 25 | { |
| 26 | struct drm_i915_private *i915 = to_i915(dev: obj->base.dev); |
| 27 | unsigned long supported = RUNTIME_INFO(i915)->page_sizes; |
| 28 | bool shrinkable; |
| 29 | int i; |
| 30 | |
| 31 | assert_object_held_shared(obj); |
| 32 | |
| 33 | if (i915_gem_object_is_volatile(obj)) |
| 34 | obj->mm.madv = I915_MADV_DONTNEED; |
| 35 | |
| 36 | /* Make the pages coherent with the GPU (flushing any swapin). */ |
| 37 | if (obj->cache_dirty) { |
| 38 | WARN_ON_ONCE(IS_DGFX(i915)); |
| 39 | obj->write_domain = 0; |
| 40 | if (i915_gem_object_has_struct_page(obj)) |
| 41 | drm_clflush_sg(st: pages); |
| 42 | obj->cache_dirty = false; |
| 43 | } |
| 44 | |
| 45 | obj->mm.get_page.sg_pos = pages->sgl; |
| 46 | obj->mm.get_page.sg_idx = 0; |
| 47 | obj->mm.get_dma_page.sg_pos = pages->sgl; |
| 48 | obj->mm.get_dma_page.sg_idx = 0; |
| 49 | |
| 50 | obj->mm.pages = pages; |
| 51 | |
| 52 | obj->mm.page_sizes.phys = i915_sg_dma_sizes(sg: pages->sgl); |
| 53 | GEM_BUG_ON(!obj->mm.page_sizes.phys); |
| 54 | |
| 55 | /* |
| 56 | * Calculate the supported page-sizes which fit into the given |
| 57 | * sg_page_sizes. This will give us the page-sizes which we may be able |
| 58 | * to use opportunistically when later inserting into the GTT. For |
| 59 | * example if phys=2G, then in theory we should be able to use 1G, 2M, |
| 60 | * 64K or 4K pages, although in practice this will depend on a number of |
| 61 | * other factors. |
| 62 | */ |
| 63 | obj->mm.page_sizes.sg = 0; |
| 64 | for_each_set_bit(i, &supported, ilog2(I915_GTT_MAX_PAGE_SIZE) + 1) { |
| 65 | if (obj->mm.page_sizes.phys & ~0u << i) |
| 66 | obj->mm.page_sizes.sg |= BIT(i); |
| 67 | } |
| 68 | GEM_BUG_ON(!HAS_PAGE_SIZES(i915, obj->mm.page_sizes.sg)); |
| 69 | |
| 70 | shrinkable = i915_gem_object_is_shrinkable(obj); |
| 71 | |
| 72 | if (i915_gem_object_is_tiled(obj) && |
| 73 | i915->gem_quirks & GEM_QUIRK_PIN_SWIZZLED_PAGES) { |
| 74 | GEM_BUG_ON(i915_gem_object_has_tiling_quirk(obj)); |
| 75 | i915_gem_object_set_tiling_quirk(obj); |
| 76 | GEM_BUG_ON(!list_empty(&obj->mm.link)); |
| 77 | atomic_inc(v: &obj->mm.shrink_pin); |
| 78 | shrinkable = false; |
| 79 | } |
| 80 | |
| 81 | if (shrinkable && !i915_gem_object_has_self_managed_shrink_list(obj)) { |
| 82 | struct list_head *list; |
| 83 | unsigned long flags; |
| 84 | |
| 85 | assert_object_held(obj); |
| 86 | spin_lock_irqsave(&i915->mm.obj_lock, flags); |
| 87 | |
| 88 | i915->mm.shrink_count++; |
| 89 | i915->mm.shrink_memory += obj->base.size; |
| 90 | |
| 91 | if (obj->mm.madv != I915_MADV_WILLNEED) |
| 92 | list = &i915->mm.purge_list; |
| 93 | else |
| 94 | list = &i915->mm.shrink_list; |
| 95 | list_add_tail(new: &obj->mm.link, head: list); |
| 96 | |
| 97 | atomic_set(v: &obj->mm.shrink_pin, i: 0); |
| 98 | spin_unlock_irqrestore(lock: &i915->mm.obj_lock, flags); |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | int ____i915_gem_object_get_pages(struct drm_i915_gem_object *obj) |
| 103 | { |
| 104 | struct drm_i915_private *i915 = to_i915(dev: obj->base.dev); |
| 105 | int err; |
| 106 | |
| 107 | assert_object_held_shared(obj); |
| 108 | |
| 109 | if (unlikely(obj->mm.madv != I915_MADV_WILLNEED)) { |
| 110 | drm_dbg(&i915->drm, |
| 111 | "Attempting to obtain a purgeable object\n" ); |
| 112 | return -EFAULT; |
| 113 | } |
| 114 | |
| 115 | err = obj->ops->get_pages(obj); |
| 116 | GEM_BUG_ON(!err && !i915_gem_object_has_pages(obj)); |
| 117 | |
| 118 | return err; |
| 119 | } |
| 120 | |
| 121 | /* Ensure that the associated pages are gathered from the backing storage |
| 122 | * and pinned into our object. i915_gem_object_pin_pages() may be called |
| 123 | * multiple times before they are released by a single call to |
| 124 | * i915_gem_object_unpin_pages() - once the pages are no longer referenced |
| 125 | * either as a result of memory pressure (reaping pages under the shrinker) |
| 126 | * or as the object is itself released. |
| 127 | */ |
| 128 | int __i915_gem_object_get_pages(struct drm_i915_gem_object *obj) |
| 129 | { |
| 130 | int err; |
| 131 | |
| 132 | assert_object_held(obj); |
| 133 | |
| 134 | assert_object_held_shared(obj); |
| 135 | |
| 136 | if (unlikely(!i915_gem_object_has_pages(obj))) { |
| 137 | GEM_BUG_ON(i915_gem_object_has_pinned_pages(obj)); |
| 138 | |
| 139 | err = ____i915_gem_object_get_pages(obj); |
| 140 | if (err) |
| 141 | return err; |
| 142 | |
| 143 | smp_mb__before_atomic(); |
| 144 | } |
| 145 | atomic_inc(v: &obj->mm.pages_pin_count); |
| 146 | |
| 147 | return 0; |
| 148 | } |
| 149 | |
| 150 | int i915_gem_object_pin_pages_unlocked(struct drm_i915_gem_object *obj) |
| 151 | { |
| 152 | struct i915_gem_ww_ctx ww; |
| 153 | int err; |
| 154 | |
| 155 | i915_gem_ww_ctx_init(ctx: &ww, intr: true); |
| 156 | retry: |
| 157 | err = i915_gem_object_lock(obj, ww: &ww); |
| 158 | if (!err) |
| 159 | err = i915_gem_object_pin_pages(obj); |
| 160 | |
| 161 | if (err == -EDEADLK) { |
| 162 | err = i915_gem_ww_ctx_backoff(ctx: &ww); |
| 163 | if (!err) |
| 164 | goto retry; |
| 165 | } |
| 166 | i915_gem_ww_ctx_fini(ctx: &ww); |
| 167 | return err; |
| 168 | } |
| 169 | |
| 170 | /* Immediately discard the backing storage */ |
| 171 | int i915_gem_object_truncate(struct drm_i915_gem_object *obj) |
| 172 | { |
| 173 | if (obj->ops->truncate) |
| 174 | return obj->ops->truncate(obj); |
| 175 | |
| 176 | return 0; |
| 177 | } |
| 178 | |
| 179 | static void __i915_gem_object_reset_page_iter(struct drm_i915_gem_object *obj) |
| 180 | { |
| 181 | struct radix_tree_iter iter; |
| 182 | void __rcu **slot; |
| 183 | |
| 184 | rcu_read_lock(); |
| 185 | radix_tree_for_each_slot(slot, &obj->mm.get_page.radix, &iter, 0) |
| 186 | radix_tree_delete(&obj->mm.get_page.radix, iter.index); |
| 187 | radix_tree_for_each_slot(slot, &obj->mm.get_dma_page.radix, &iter, 0) |
| 188 | radix_tree_delete(&obj->mm.get_dma_page.radix, iter.index); |
| 189 | rcu_read_unlock(); |
| 190 | } |
| 191 | |
| 192 | static void unmap_object(struct drm_i915_gem_object *obj, void *ptr) |
| 193 | { |
| 194 | if (is_vmalloc_addr(x: ptr)) |
| 195 | vunmap(addr: ptr); |
| 196 | } |
| 197 | |
| 198 | static void flush_tlb_invalidate(struct drm_i915_gem_object *obj) |
| 199 | { |
| 200 | struct drm_i915_private *i915 = to_i915(dev: obj->base.dev); |
| 201 | struct intel_gt *gt; |
| 202 | int id; |
| 203 | |
| 204 | for_each_gt(gt, i915, id) { |
| 205 | if (!obj->mm.tlb[id]) |
| 206 | continue; |
| 207 | |
| 208 | intel_gt_invalidate_tlb_full(gt, seqno: obj->mm.tlb[id]); |
| 209 | obj->mm.tlb[id] = 0; |
| 210 | } |
| 211 | } |
| 212 | |
| 213 | struct sg_table * |
| 214 | __i915_gem_object_unset_pages(struct drm_i915_gem_object *obj) |
| 215 | { |
| 216 | struct sg_table *pages; |
| 217 | |
| 218 | assert_object_held_shared(obj); |
| 219 | |
| 220 | pages = fetch_and_zero(&obj->mm.pages); |
| 221 | if (IS_ERR_OR_NULL(ptr: pages)) |
| 222 | return pages; |
| 223 | |
| 224 | if (i915_gem_object_is_volatile(obj)) |
| 225 | obj->mm.madv = I915_MADV_WILLNEED; |
| 226 | |
| 227 | if (!i915_gem_object_has_self_managed_shrink_list(obj)) |
| 228 | i915_gem_object_make_unshrinkable(obj); |
| 229 | |
| 230 | if (obj->mm.mapping) { |
| 231 | unmap_object(obj, page_mask_bits(obj->mm.mapping)); |
| 232 | obj->mm.mapping = NULL; |
| 233 | } |
| 234 | |
| 235 | __i915_gem_object_reset_page_iter(obj); |
| 236 | obj->mm.page_sizes.phys = obj->mm.page_sizes.sg = 0; |
| 237 | |
| 238 | flush_tlb_invalidate(obj); |
| 239 | |
| 240 | return pages; |
| 241 | } |
| 242 | |
| 243 | int __i915_gem_object_put_pages(struct drm_i915_gem_object *obj) |
| 244 | { |
| 245 | struct sg_table *pages; |
| 246 | |
| 247 | if (i915_gem_object_has_pinned_pages(obj)) |
| 248 | return -EBUSY; |
| 249 | |
| 250 | /* May be called by shrinker from within get_pages() (on another bo) */ |
| 251 | assert_object_held_shared(obj); |
| 252 | |
| 253 | i915_gem_object_release_mmap_offset(obj); |
| 254 | |
| 255 | /* |
| 256 | * ->put_pages might need to allocate memory for the bit17 swizzle |
| 257 | * array, hence protect them from being reaped by removing them from gtt |
| 258 | * lists early. |
| 259 | */ |
| 260 | pages = __i915_gem_object_unset_pages(obj); |
| 261 | |
| 262 | /* |
| 263 | * XXX Temporary hijinx to avoid updating all backends to handle |
| 264 | * NULL pages. In the future, when we have more asynchronous |
| 265 | * get_pages backends we should be better able to handle the |
| 266 | * cancellation of the async task in a more uniform manner. |
| 267 | */ |
| 268 | if (!IS_ERR_OR_NULL(ptr: pages)) |
| 269 | obj->ops->put_pages(obj, pages); |
| 270 | |
| 271 | return 0; |
| 272 | } |
| 273 | |
| 274 | /* The 'mapping' part of i915_gem_object_pin_map() below */ |
| 275 | static void *i915_gem_object_map_page(struct drm_i915_gem_object *obj, |
| 276 | enum i915_map_type type) |
| 277 | { |
| 278 | unsigned long n_pages = obj->base.size >> PAGE_SHIFT, i; |
| 279 | struct page *stack[32], **pages = stack, *page; |
| 280 | struct sgt_iter iter; |
| 281 | pgprot_t pgprot; |
| 282 | void *vaddr; |
| 283 | |
| 284 | switch (type) { |
| 285 | default: |
| 286 | MISSING_CASE(type); |
| 287 | fallthrough; /* to use PAGE_KERNEL anyway */ |
| 288 | case I915_MAP_WB: |
| 289 | /* |
| 290 | * On 32b, highmem using a finite set of indirect PTE (i.e. |
| 291 | * vmap) to provide virtual mappings of the high pages. |
| 292 | * As these are finite, map_new_virtual() must wait for some |
| 293 | * other kmap() to finish when it runs out. If we map a large |
| 294 | * number of objects, there is no method for it to tell us |
| 295 | * to release the mappings, and we deadlock. |
| 296 | * |
| 297 | * However, if we make an explicit vmap of the page, that |
| 298 | * uses a larger vmalloc arena, and also has the ability |
| 299 | * to tell us to release unwanted mappings. Most importantly, |
| 300 | * it will fail and propagate an error instead of waiting |
| 301 | * forever. |
| 302 | * |
| 303 | * So if the page is beyond the 32b boundary, make an explicit |
| 304 | * vmap. |
| 305 | */ |
| 306 | if (n_pages == 1 && !PageHighMem(page: sg_page(sg: obj->mm.pages->sgl))) |
| 307 | return page_address(sg_page(obj->mm.pages->sgl)); |
| 308 | pgprot = PAGE_KERNEL; |
| 309 | break; |
| 310 | case I915_MAP_WC: |
| 311 | pgprot = pgprot_writecombine(PAGE_KERNEL_IO); |
| 312 | break; |
| 313 | } |
| 314 | |
| 315 | if (n_pages > ARRAY_SIZE(stack)) { |
| 316 | /* Too big for stack -- allocate temporary array instead */ |
| 317 | pages = kvmalloc_array(n_pages, sizeof(*pages), GFP_KERNEL); |
| 318 | if (!pages) |
| 319 | return ERR_PTR(error: -ENOMEM); |
| 320 | } |
| 321 | |
| 322 | i = 0; |
| 323 | for_each_sgt_page(page, iter, obj->mm.pages) |
| 324 | pages[i++] = page; |
| 325 | vaddr = vmap(pages, count: n_pages, flags: 0, prot: pgprot); |
| 326 | if (pages != stack) |
| 327 | kvfree(addr: pages); |
| 328 | |
| 329 | return vaddr ?: ERR_PTR(error: -ENOMEM); |
| 330 | } |
| 331 | |
| 332 | static void *i915_gem_object_map_pfn(struct drm_i915_gem_object *obj, |
| 333 | enum i915_map_type type) |
| 334 | { |
| 335 | resource_size_t iomap = obj->mm.region->iomap.base - |
| 336 | obj->mm.region->region.start; |
| 337 | unsigned long n_pfn = obj->base.size >> PAGE_SHIFT; |
| 338 | unsigned long stack[32], *pfns = stack, i; |
| 339 | struct sgt_iter iter; |
| 340 | dma_addr_t addr; |
| 341 | void *vaddr; |
| 342 | |
| 343 | GEM_BUG_ON(type != I915_MAP_WC); |
| 344 | |
| 345 | if (n_pfn > ARRAY_SIZE(stack)) { |
| 346 | /* Too big for stack -- allocate temporary array instead */ |
| 347 | pfns = kvmalloc_array(n_pfn, sizeof(*pfns), GFP_KERNEL); |
| 348 | if (!pfns) |
| 349 | return ERR_PTR(error: -ENOMEM); |
| 350 | } |
| 351 | |
| 352 | i = 0; |
| 353 | for_each_sgt_daddr(addr, iter, obj->mm.pages) |
| 354 | pfns[i++] = (iomap + addr) >> PAGE_SHIFT; |
| 355 | vaddr = vmap_pfn(pfns, count: n_pfn, pgprot_writecombine(PAGE_KERNEL_IO)); |
| 356 | if (pfns != stack) |
| 357 | kvfree(addr: pfns); |
| 358 | |
| 359 | return vaddr ?: ERR_PTR(error: -ENOMEM); |
| 360 | } |
| 361 | |
| 362 | struct intel_panic { |
| 363 | struct page **pages; |
| 364 | int page; |
| 365 | void *vaddr; |
| 366 | }; |
| 367 | |
| 368 | static void i915_panic_kunmap(struct intel_panic *panic) |
| 369 | { |
| 370 | if (panic->vaddr) { |
| 371 | drm_clflush_virt_range(addr: panic->vaddr, PAGE_SIZE); |
| 372 | kunmap_local(panic->vaddr); |
| 373 | panic->vaddr = NULL; |
| 374 | } |
| 375 | } |
| 376 | |
| 377 | static struct page **i915_gem_object_panic_pages(struct drm_i915_gem_object *obj) |
| 378 | { |
| 379 | unsigned long n_pages = obj->base.size >> PAGE_SHIFT, i; |
| 380 | struct page *page; |
| 381 | struct page **pages; |
| 382 | struct sgt_iter iter; |
| 383 | |
| 384 | /* For a 3840x2160 32 bits Framebuffer, this should require ~64K */ |
| 385 | pages = kmalloc_array(n_pages, sizeof(*pages), GFP_ATOMIC); |
| 386 | if (!pages) |
| 387 | return NULL; |
| 388 | |
| 389 | i = 0; |
| 390 | for_each_sgt_page(page, iter, obj->mm.pages) |
| 391 | pages[i++] = page; |
| 392 | return pages; |
| 393 | } |
| 394 | |
| 395 | static void i915_gem_object_panic_map_set_pixel(struct drm_scanout_buffer *sb, unsigned int x, |
| 396 | unsigned int y, u32 color) |
| 397 | { |
| 398 | struct intel_framebuffer *fb = (struct intel_framebuffer *)sb->private; |
| 399 | unsigned int offset = fb->panic_tiling(sb->width, x, y); |
| 400 | |
| 401 | iosys_map_wr(&sb->map[0], offset, u32, color); |
| 402 | } |
| 403 | |
| 404 | /* |
| 405 | * The scanout buffer pages are not mapped, so for each pixel, |
| 406 | * use kmap_local_page_try_from_panic() to map the page, and write the pixel. |
| 407 | * Try to keep the map from the previous pixel, to avoid too much map/unmap. |
| 408 | */ |
| 409 | static void i915_gem_object_panic_page_set_pixel(struct drm_scanout_buffer *sb, unsigned int x, |
| 410 | unsigned int y, u32 color) |
| 411 | { |
| 412 | unsigned int new_page; |
| 413 | unsigned int offset; |
| 414 | struct intel_framebuffer *fb = (struct intel_framebuffer *)sb->private; |
| 415 | struct intel_panic *panic = fb->panic; |
| 416 | |
| 417 | if (fb->panic_tiling) |
| 418 | offset = fb->panic_tiling(sb->width, x, y); |
| 419 | else |
| 420 | offset = y * sb->pitch[0] + x * sb->format->cpp[0]; |
| 421 | |
| 422 | new_page = offset >> PAGE_SHIFT; |
| 423 | offset = offset % PAGE_SIZE; |
| 424 | if (new_page != panic->page) { |
| 425 | i915_panic_kunmap(panic); |
| 426 | panic->page = new_page; |
| 427 | panic->vaddr = |
| 428 | kmap_local_page_try_from_panic(page: panic->pages[panic->page]); |
| 429 | } |
| 430 | if (panic->vaddr) { |
| 431 | u32 *pix = panic->vaddr + offset; |
| 432 | *pix = color; |
| 433 | } |
| 434 | } |
| 435 | |
| 436 | struct intel_panic *i915_gem_object_alloc_panic(void) |
| 437 | { |
| 438 | struct intel_panic *panic; |
| 439 | |
| 440 | panic = kzalloc(sizeof(*panic), GFP_KERNEL); |
| 441 | |
| 442 | return panic; |
| 443 | } |
| 444 | |
| 445 | /* |
| 446 | * Setup the gem framebuffer for drm_panic access. |
| 447 | * Use current vaddr if it exists, or setup a list of pages. |
| 448 | * pfn is not supported yet. |
| 449 | */ |
| 450 | int i915_gem_object_panic_setup(struct intel_panic *panic, struct drm_scanout_buffer *sb, |
| 451 | struct drm_gem_object *_obj, bool panic_tiling) |
| 452 | { |
| 453 | enum i915_map_type has_type; |
| 454 | struct drm_i915_gem_object *obj = to_intel_bo(gem: _obj); |
| 455 | void *ptr; |
| 456 | |
| 457 | ptr = page_unpack_bits(obj->mm.mapping, &has_type); |
| 458 | if (ptr) { |
| 459 | if (i915_gem_object_has_iomem(obj)) |
| 460 | iosys_map_set_vaddr_iomem(map: &sb->map[0], vaddr_iomem: (void __iomem *)ptr); |
| 461 | else |
| 462 | iosys_map_set_vaddr(map: &sb->map[0], vaddr: ptr); |
| 463 | |
| 464 | if (panic_tiling) |
| 465 | sb->set_pixel = i915_gem_object_panic_map_set_pixel; |
| 466 | return 0; |
| 467 | } |
| 468 | if (i915_gem_object_has_struct_page(obj)) { |
| 469 | panic->pages = i915_gem_object_panic_pages(obj); |
| 470 | if (!panic->pages) |
| 471 | return -ENOMEM; |
| 472 | panic->page = -1; |
| 473 | sb->set_pixel = i915_gem_object_panic_page_set_pixel; |
| 474 | return 0; |
| 475 | } |
| 476 | return -EOPNOTSUPP; |
| 477 | } |
| 478 | |
| 479 | void i915_gem_object_panic_finish(struct intel_panic *panic) |
| 480 | { |
| 481 | i915_panic_kunmap(panic); |
| 482 | panic->page = -1; |
| 483 | kfree(objp: panic->pages); |
| 484 | panic->pages = NULL; |
| 485 | } |
| 486 | |
| 487 | /* get, pin, and map the pages of the object into kernel space */ |
| 488 | void *i915_gem_object_pin_map(struct drm_i915_gem_object *obj, |
| 489 | enum i915_map_type type) |
| 490 | { |
| 491 | enum i915_map_type has_type; |
| 492 | bool pinned; |
| 493 | void *ptr; |
| 494 | int err; |
| 495 | |
| 496 | if (!i915_gem_object_has_struct_page(obj) && |
| 497 | !i915_gem_object_has_iomem(obj)) |
| 498 | return ERR_PTR(error: -ENXIO); |
| 499 | |
| 500 | if (WARN_ON_ONCE(obj->flags & I915_BO_ALLOC_GPU_ONLY)) |
| 501 | return ERR_PTR(error: -EINVAL); |
| 502 | |
| 503 | assert_object_held(obj); |
| 504 | |
| 505 | pinned = !(type & I915_MAP_OVERRIDE); |
| 506 | type &= ~I915_MAP_OVERRIDE; |
| 507 | |
| 508 | if (!atomic_inc_not_zero(v: &obj->mm.pages_pin_count)) { |
| 509 | if (unlikely(!i915_gem_object_has_pages(obj))) { |
| 510 | GEM_BUG_ON(i915_gem_object_has_pinned_pages(obj)); |
| 511 | |
| 512 | err = ____i915_gem_object_get_pages(obj); |
| 513 | if (err) |
| 514 | return ERR_PTR(error: err); |
| 515 | |
| 516 | smp_mb__before_atomic(); |
| 517 | } |
| 518 | atomic_inc(v: &obj->mm.pages_pin_count); |
| 519 | pinned = false; |
| 520 | } |
| 521 | GEM_BUG_ON(!i915_gem_object_has_pages(obj)); |
| 522 | |
| 523 | /* |
| 524 | * For discrete our CPU mappings needs to be consistent in order to |
| 525 | * function correctly on !x86. When mapping things through TTM, we use |
| 526 | * the same rules to determine the caching type. |
| 527 | * |
| 528 | * The caching rules, starting from DG1: |
| 529 | * |
| 530 | * - If the object can be placed in device local-memory, then the |
| 531 | * pages should be allocated and mapped as write-combined only. |
| 532 | * |
| 533 | * - Everything else is always allocated and mapped as write-back, |
| 534 | * with the guarantee that everything is also coherent with the |
| 535 | * GPU. |
| 536 | * |
| 537 | * Internal users of lmem are already expected to get this right, so no |
| 538 | * fudging needed there. |
| 539 | */ |
| 540 | if (i915_gem_object_placement_possible(obj, type: INTEL_MEMORY_LOCAL)) { |
| 541 | if (type != I915_MAP_WC && !obj->mm.n_placements) { |
| 542 | ptr = ERR_PTR(error: -ENODEV); |
| 543 | goto err_unpin; |
| 544 | } |
| 545 | |
| 546 | type = I915_MAP_WC; |
| 547 | } else if (IS_DGFX(to_i915(obj->base.dev))) { |
| 548 | type = I915_MAP_WB; |
| 549 | } |
| 550 | |
| 551 | ptr = page_unpack_bits(obj->mm.mapping, &has_type); |
| 552 | if (ptr && has_type != type) { |
| 553 | if (pinned) { |
| 554 | ptr = ERR_PTR(error: -EBUSY); |
| 555 | goto err_unpin; |
| 556 | } |
| 557 | |
| 558 | unmap_object(obj, ptr); |
| 559 | |
| 560 | ptr = obj->mm.mapping = NULL; |
| 561 | } |
| 562 | |
| 563 | if (!ptr) { |
| 564 | err = i915_gem_object_wait_moving_fence(obj, intr: true); |
| 565 | if (err) { |
| 566 | ptr = ERR_PTR(error: err); |
| 567 | goto err_unpin; |
| 568 | } |
| 569 | |
| 570 | if (GEM_WARN_ON(type == I915_MAP_WC && !pat_enabled())) |
| 571 | ptr = ERR_PTR(error: -ENODEV); |
| 572 | else if (i915_gem_object_has_struct_page(obj)) |
| 573 | ptr = i915_gem_object_map_page(obj, type); |
| 574 | else |
| 575 | ptr = i915_gem_object_map_pfn(obj, type); |
| 576 | if (IS_ERR(ptr)) |
| 577 | goto err_unpin; |
| 578 | |
| 579 | obj->mm.mapping = page_pack_bits(ptr, type); |
| 580 | } |
| 581 | |
| 582 | return ptr; |
| 583 | |
| 584 | err_unpin: |
| 585 | atomic_dec(v: &obj->mm.pages_pin_count); |
| 586 | return ptr; |
| 587 | } |
| 588 | |
| 589 | void *i915_gem_object_pin_map_unlocked(struct drm_i915_gem_object *obj, |
| 590 | enum i915_map_type type) |
| 591 | { |
| 592 | void *ret; |
| 593 | |
| 594 | i915_gem_object_lock(obj, NULL); |
| 595 | ret = i915_gem_object_pin_map(obj, type); |
| 596 | i915_gem_object_unlock(obj); |
| 597 | |
| 598 | return ret; |
| 599 | } |
| 600 | |
| 601 | void __i915_gem_object_flush_map(struct drm_i915_gem_object *obj, |
| 602 | unsigned long offset, |
| 603 | unsigned long size) |
| 604 | { |
| 605 | enum i915_map_type has_type; |
| 606 | void *ptr; |
| 607 | |
| 608 | GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj)); |
| 609 | GEM_BUG_ON(range_overflows_t(typeof(obj->base.size), |
| 610 | offset, size, obj->base.size)); |
| 611 | |
| 612 | wmb(); /* let all previous writes be visible to coherent partners */ |
| 613 | obj->mm.dirty = true; |
| 614 | |
| 615 | if (obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE) |
| 616 | return; |
| 617 | |
| 618 | ptr = page_unpack_bits(obj->mm.mapping, &has_type); |
| 619 | if (has_type == I915_MAP_WC) |
| 620 | return; |
| 621 | |
| 622 | drm_clflush_virt_range(addr: ptr + offset, length: size); |
| 623 | if (size == obj->base.size) { |
| 624 | obj->write_domain &= ~I915_GEM_DOMAIN_CPU; |
| 625 | obj->cache_dirty = false; |
| 626 | } |
| 627 | } |
| 628 | |
| 629 | void __i915_gem_object_release_map(struct drm_i915_gem_object *obj) |
| 630 | { |
| 631 | GEM_BUG_ON(!obj->mm.mapping); |
| 632 | |
| 633 | /* |
| 634 | * We allow removing the mapping from underneath pinned pages! |
| 635 | * |
| 636 | * Furthermore, since this is an unsafe operation reserved only |
| 637 | * for construction time manipulation, we ignore locking prudence. |
| 638 | */ |
| 639 | unmap_object(obj, page_mask_bits(fetch_and_zero(&obj->mm.mapping))); |
| 640 | |
| 641 | i915_gem_object_unpin_map(obj); |
| 642 | } |
| 643 | |
| 644 | struct scatterlist * |
| 645 | __i915_gem_object_page_iter_get_sg(struct drm_i915_gem_object *obj, |
| 646 | struct i915_gem_object_page_iter *iter, |
| 647 | pgoff_t n, |
| 648 | unsigned int *offset) |
| 649 | |
| 650 | { |
| 651 | const bool dma = iter == &obj->mm.get_dma_page || |
| 652 | iter == &obj->ttm.get_io_page; |
| 653 | unsigned int idx, count; |
| 654 | struct scatterlist *sg; |
| 655 | |
| 656 | might_sleep(); |
| 657 | GEM_BUG_ON(n >= obj->base.size >> PAGE_SHIFT); |
| 658 | if (!i915_gem_object_has_pinned_pages(obj)) |
| 659 | assert_object_held(obj); |
| 660 | |
| 661 | /* As we iterate forward through the sg, we record each entry in a |
| 662 | * radixtree for quick repeated (backwards) lookups. If we have seen |
| 663 | * this index previously, we will have an entry for it. |
| 664 | * |
| 665 | * Initial lookup is O(N), but this is amortized to O(1) for |
| 666 | * sequential page access (where each new request is consecutive |
| 667 | * to the previous one). Repeated lookups are O(lg(obj->base.size)), |
| 668 | * i.e. O(1) with a large constant! |
| 669 | */ |
| 670 | if (n < READ_ONCE(iter->sg_idx)) |
| 671 | goto lookup; |
| 672 | |
| 673 | mutex_lock(&iter->lock); |
| 674 | |
| 675 | /* We prefer to reuse the last sg so that repeated lookup of this |
| 676 | * (or the subsequent) sg are fast - comparing against the last |
| 677 | * sg is faster than going through the radixtree. |
| 678 | */ |
| 679 | |
| 680 | sg = iter->sg_pos; |
| 681 | idx = iter->sg_idx; |
| 682 | count = dma ? __sg_dma_page_count(sg) : __sg_page_count(sg); |
| 683 | |
| 684 | while (idx + count <= n) { |
| 685 | void *entry; |
| 686 | unsigned long i; |
| 687 | int ret; |
| 688 | |
| 689 | /* If we cannot allocate and insert this entry, or the |
| 690 | * individual pages from this range, cancel updating the |
| 691 | * sg_idx so that on this lookup we are forced to linearly |
| 692 | * scan onwards, but on future lookups we will try the |
| 693 | * insertion again (in which case we need to be careful of |
| 694 | * the error return reporting that we have already inserted |
| 695 | * this index). |
| 696 | */ |
| 697 | ret = radix_tree_insert(&iter->radix, index: idx, sg); |
| 698 | if (ret && ret != -EEXIST) |
| 699 | goto scan; |
| 700 | |
| 701 | entry = xa_mk_value(v: idx); |
| 702 | for (i = 1; i < count; i++) { |
| 703 | ret = radix_tree_insert(&iter->radix, index: idx + i, entry); |
| 704 | if (ret && ret != -EEXIST) |
| 705 | goto scan; |
| 706 | } |
| 707 | |
| 708 | idx += count; |
| 709 | sg = ____sg_next(sg); |
| 710 | count = dma ? __sg_dma_page_count(sg) : __sg_page_count(sg); |
| 711 | } |
| 712 | |
| 713 | scan: |
| 714 | iter->sg_pos = sg; |
| 715 | iter->sg_idx = idx; |
| 716 | |
| 717 | mutex_unlock(lock: &iter->lock); |
| 718 | |
| 719 | if (unlikely(n < idx)) /* insertion completed by another thread */ |
| 720 | goto lookup; |
| 721 | |
| 722 | /* In case we failed to insert the entry into the radixtree, we need |
| 723 | * to look beyond the current sg. |
| 724 | */ |
| 725 | while (idx + count <= n) { |
| 726 | idx += count; |
| 727 | sg = ____sg_next(sg); |
| 728 | count = dma ? __sg_dma_page_count(sg) : __sg_page_count(sg); |
| 729 | } |
| 730 | |
| 731 | *offset = n - idx; |
| 732 | return sg; |
| 733 | |
| 734 | lookup: |
| 735 | rcu_read_lock(); |
| 736 | |
| 737 | sg = radix_tree_lookup(&iter->radix, n); |
| 738 | GEM_BUG_ON(!sg); |
| 739 | |
| 740 | /* If this index is in the middle of multi-page sg entry, |
| 741 | * the radix tree will contain a value entry that points |
| 742 | * to the start of that range. We will return the pointer to |
| 743 | * the base page and the offset of this page within the |
| 744 | * sg entry's range. |
| 745 | */ |
| 746 | *offset = 0; |
| 747 | if (unlikely(xa_is_value(sg))) { |
| 748 | unsigned long base = xa_to_value(entry: sg); |
| 749 | |
| 750 | sg = radix_tree_lookup(&iter->radix, base); |
| 751 | GEM_BUG_ON(!sg); |
| 752 | |
| 753 | *offset = n - base; |
| 754 | } |
| 755 | |
| 756 | rcu_read_unlock(); |
| 757 | |
| 758 | return sg; |
| 759 | } |
| 760 | |
| 761 | struct page * |
| 762 | __i915_gem_object_get_page(struct drm_i915_gem_object *obj, pgoff_t n) |
| 763 | { |
| 764 | struct scatterlist *sg; |
| 765 | unsigned int offset; |
| 766 | |
| 767 | GEM_BUG_ON(!i915_gem_object_has_struct_page(obj)); |
| 768 | |
| 769 | sg = i915_gem_object_get_sg(obj, n, &offset); |
| 770 | return sg_page(sg) + offset; |
| 771 | } |
| 772 | |
| 773 | /* Like i915_gem_object_get_page(), but mark the returned page dirty */ |
| 774 | struct page * |
| 775 | __i915_gem_object_get_dirty_page(struct drm_i915_gem_object *obj, pgoff_t n) |
| 776 | { |
| 777 | struct page *page; |
| 778 | |
| 779 | page = i915_gem_object_get_page(obj, n); |
| 780 | if (!obj->mm.dirty) |
| 781 | set_page_dirty(page); |
| 782 | |
| 783 | return page; |
| 784 | } |
| 785 | |
| 786 | dma_addr_t |
| 787 | __i915_gem_object_get_dma_address_len(struct drm_i915_gem_object *obj, |
| 788 | pgoff_t n, unsigned int *len) |
| 789 | { |
| 790 | struct scatterlist *sg; |
| 791 | unsigned int offset; |
| 792 | |
| 793 | sg = i915_gem_object_get_sg_dma(obj, n, &offset); |
| 794 | |
| 795 | if (len) |
| 796 | *len = sg_dma_len(sg) - (offset << PAGE_SHIFT); |
| 797 | |
| 798 | return sg_dma_address(sg) + (offset << PAGE_SHIFT); |
| 799 | } |
| 800 | |
| 801 | dma_addr_t |
| 802 | __i915_gem_object_get_dma_address(struct drm_i915_gem_object *obj, pgoff_t n) |
| 803 | { |
| 804 | return i915_gem_object_get_dma_address_len(obj, n, NULL); |
| 805 | } |
| 806 | |