1/* $NetBSD: i915_gem_fence_reg.c,v 1.6 2021/12/19 11:33:49 riastradh Exp $ */ 2 3/* 4 * Copyright �� 2008-2015 Intel Corporation 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a 7 * copy of this software and associated documentation files (the "Software"), 8 * to deal in the Software without restriction, including without limitation 9 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 10 * and/or sell copies of the Software, and to permit persons to whom the 11 * Software is furnished to do so, subject to the following conditions: 12 * 13 * The above copyright notice and this permission notice (including the next 14 * paragraph) shall be included in all copies or substantial portions of the 15 * Software. 16 * 17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 22 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS 23 * IN THE SOFTWARE. 24 */ 25 26#include <sys/cdefs.h> 27__KERNEL_RCSID(0, "$NetBSD: i915_gem_fence_reg.c,v 1.6 2021/12/19 11:33:49 riastradh Exp $"); 28 29#include <linux/bitmap.h> 30#include <drm/i915_drm.h> 31 32#include "i915_drv.h" 33#include "i915_scatterlist.h" 34#include "i915_vgpu.h" 35 36/** 37 * DOC: fence register handling 38 * 39 * Important to avoid confusions: "fences" in the i915 driver are not execution 40 * fences used to track command completion but hardware detiler objects which 41 * wrap a given range of the global GTT. Each platform has only a fairly limited 42 * set of these objects. 43 * 44 * Fences are used to detile GTT memory mappings. They're also connected to the 45 * hardware frontbuffer render tracking and hence interact with frontbuffer 46 * compression. Furthermore on older platforms fences are required for tiled 47 * objects used by the display engine. They can also be used by the render 48 * engine - they're required for blitter commands and are optional for render 49 * commands. But on gen4+ both display (with the exception of fbc) and rendering 50 * have their own tiling state bits and don't need fences. 51 * 52 * Also note that fences only support X and Y tiling and hence can't be used for 53 * the fancier new tiling formats like W, Ys and Yf. 54 * 55 * Finally note that because fences are such a restricted resource they're 56 * dynamically associated with objects. Furthermore fence state is committed to 57 * the hardware lazily to avoid unnecessary stalls on gen2/3. Therefore code must 58 * explicitly call i915_gem_object_get_fence() to synchronize fencing status 59 * for cpu access. Also note that some code wants an unfenced view, for those 60 * cases the fence can be removed forcefully with i915_gem_object_put_fence(). 61 * 62 * Internally these functions will synchronize with userspace access by removing 63 * CPU ptes into GTT mmaps (not the GTT ptes themselves) as needed. 64 */ 65 66#define pipelined 0 67 68static struct drm_i915_private *fence_to_i915(struct i915_fence_reg *fence) 69{ 70 return fence->ggtt->vm.i915; 71} 72 73static struct intel_uncore *fence_to_uncore(struct i915_fence_reg *fence) 74{ 75 return fence->ggtt->vm.gt->uncore; 76} 77 78static void i965_write_fence_reg(struct i915_fence_reg *fence, 79 struct i915_vma *vma) 80{ 81 i915_reg_t fence_reg_lo, fence_reg_hi; 82 int fence_pitch_shift; 83 u64 val; 84 85 if (INTEL_GEN(fence_to_i915(fence)) >= 6) { 86 fence_reg_lo = FENCE_REG_GEN6_LO(fence->id); 87 fence_reg_hi = FENCE_REG_GEN6_HI(fence->id); 88 fence_pitch_shift = GEN6_FENCE_PITCH_SHIFT; 89 90 } else { 91 fence_reg_lo = FENCE_REG_965_LO(fence->id); 92 fence_reg_hi = FENCE_REG_965_HI(fence->id); 93 fence_pitch_shift = I965_FENCE_PITCH_SHIFT; 94 } 95 96 val = 0; 97 if (vma) { 98 unsigned int stride = i915_gem_object_get_stride(vma->obj); 99 100 GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma)); 101 GEM_BUG_ON(!IS_ALIGNED(vma->node.start, I965_FENCE_PAGE)); 102 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I965_FENCE_PAGE)); 103 GEM_BUG_ON(!IS_ALIGNED(stride, 128)); 104 105 val = (vma->node.start + vma->fence_size - I965_FENCE_PAGE) << 32; 106 val |= vma->node.start; 107 val |= (u64)((stride / 128) - 1) << fence_pitch_shift; 108 if (i915_gem_object_get_tiling(vma->obj) == I915_TILING_Y) 109 val |= BIT(I965_FENCE_TILING_Y_SHIFT); 110 val |= I965_FENCE_REG_VALID; 111 } 112 113 if (!pipelined) { 114 struct intel_uncore *uncore = fence_to_uncore(fence); 115 116 /* 117 * To w/a incoherency with non-atomic 64-bit register updates, 118 * we split the 64-bit update into two 32-bit writes. In order 119 * for a partial fence not to be evaluated between writes, we 120 * precede the update with write to turn off the fence register, 121 * and only enable the fence as the last step. 122 * 123 * For extra levels of paranoia, we make sure each step lands 124 * before applying the next step. 125 */ 126 intel_uncore_write_fw(uncore, fence_reg_lo, 0); 127 intel_uncore_posting_read_fw(uncore, fence_reg_lo); 128 129 intel_uncore_write_fw(uncore, fence_reg_hi, upper_32_bits(val)); 130 intel_uncore_write_fw(uncore, fence_reg_lo, lower_32_bits(val)); 131 intel_uncore_posting_read_fw(uncore, fence_reg_lo); 132 } 133} 134 135static void i915_write_fence_reg(struct i915_fence_reg *fence, 136 struct i915_vma *vma) 137{ 138 u32 val; 139 140 val = 0; 141 if (vma) { 142 unsigned int tiling = i915_gem_object_get_tiling(vma->obj); 143 bool is_y_tiled = tiling == I915_TILING_Y; 144 unsigned int stride = i915_gem_object_get_stride(vma->obj); 145 146 GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma)); 147 GEM_BUG_ON(vma->node.start & ~I915_FENCE_START_MASK); 148 GEM_BUG_ON(!is_power_of_2(vma->fence_size)); 149 GEM_BUG_ON(!IS_ALIGNED(vma->node.start, vma->fence_size)); 150 151 if (is_y_tiled && HAS_128_BYTE_Y_TILING(fence_to_i915(fence))) 152 stride /= 128; 153 else 154 stride /= 512; 155 GEM_BUG_ON(!is_power_of_2(stride)); 156 157 val = vma->node.start; 158 if (is_y_tiled) 159 val |= BIT(I830_FENCE_TILING_Y_SHIFT); 160 val |= I915_FENCE_SIZE_BITS(vma->fence_size); 161 val |= ilog2(stride) << I830_FENCE_PITCH_SHIFT; 162 163 val |= I830_FENCE_REG_VALID; 164 } 165 166 if (!pipelined) { 167 struct intel_uncore *uncore = fence_to_uncore(fence); 168 i915_reg_t reg = FENCE_REG(fence->id); 169 170 intel_uncore_write_fw(uncore, reg, val); 171 intel_uncore_posting_read_fw(uncore, reg); 172 } 173} 174 175static void i830_write_fence_reg(struct i915_fence_reg *fence, 176 struct i915_vma *vma) 177{ 178 u32 val; 179 180 val = 0; 181 if (vma) { 182 unsigned int stride = i915_gem_object_get_stride(vma->obj); 183 184 GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma)); 185 GEM_BUG_ON(vma->node.start & ~I830_FENCE_START_MASK); 186 GEM_BUG_ON(!is_power_of_2(vma->fence_size)); 187 GEM_BUG_ON(!is_power_of_2(stride / 128)); 188 GEM_BUG_ON(!IS_ALIGNED(vma->node.start, vma->fence_size)); 189 190 val = vma->node.start; 191 if (i915_gem_object_get_tiling(vma->obj) == I915_TILING_Y) 192 val |= BIT(I830_FENCE_TILING_Y_SHIFT); 193 val |= I830_FENCE_SIZE_BITS(vma->fence_size); 194 val |= ilog2(stride / 128) << I830_FENCE_PITCH_SHIFT; 195 val |= I830_FENCE_REG_VALID; 196 } 197 198 if (!pipelined) { 199 struct intel_uncore *uncore = fence_to_uncore(fence); 200 i915_reg_t reg = FENCE_REG(fence->id); 201 202 intel_uncore_write_fw(uncore, reg, val); 203 intel_uncore_posting_read_fw(uncore, reg); 204 } 205} 206 207static void fence_write(struct i915_fence_reg *fence, 208 struct i915_vma *vma) 209{ 210 struct drm_i915_private *i915 = fence_to_i915(fence); 211 212 /* 213 * Previous access through the fence register is marshalled by 214 * the mb() inside the fault handlers (i915_gem_release_mmaps) 215 * and explicitly managed for internal users. 216 */ 217 218 if (IS_GEN(i915, 2)) 219 i830_write_fence_reg(fence, vma); 220 else if (IS_GEN(i915, 3)) 221 i915_write_fence_reg(fence, vma); 222 else 223 i965_write_fence_reg(fence, vma); 224 225 /* 226 * Access through the fenced region afterwards is 227 * ordered by the posting reads whilst writing the registers. 228 */ 229 230 fence->dirty = false; 231} 232 233static int fence_update(struct i915_fence_reg *fence, 234 struct i915_vma *vma) 235{ 236 struct i915_ggtt *ggtt = fence->ggtt; 237 struct intel_uncore *uncore = fence_to_uncore(fence); 238 intel_wakeref_t wakeref; 239 struct i915_vma *old; 240 int ret; 241 242 if (vma) { 243 if (!i915_vma_is_map_and_fenceable(vma)) 244 return -EINVAL; 245 246 if (WARN(!i915_gem_object_get_stride(vma->obj) || 247 !i915_gem_object_get_tiling(vma->obj), 248 "bogus fence setup with stride: 0x%x, tiling mode: %i\n", 249 i915_gem_object_get_stride(vma->obj), 250 i915_gem_object_get_tiling(vma->obj))) 251 return -EINVAL; 252 253 ret = i915_vma_sync(vma); 254 if (ret) 255 return ret; 256 } 257 258 old = xchg(&fence->vma, NULL); 259 if (old) { 260 /* XXX Ideally we would move the waiting to outside the mutex */ 261 ret = i915_vma_sync(old); 262 if (ret) { 263 fence->vma = old; 264 return ret; 265 } 266 267 i915_vma_flush_writes(old); 268 269 /* 270 * Ensure that all userspace CPU access is completed before 271 * stealing the fence. 272 */ 273 if (old != vma) { 274 GEM_BUG_ON(old->fence != fence); 275 i915_vma_revoke_mmap(old); 276 old->fence = NULL; 277 } 278 279 list_move(&fence->link, &ggtt->fence_list); 280 } 281 282 /* 283 * We only need to update the register itself if the device is awake. 284 * If the device is currently powered down, we will defer the write 285 * to the runtime resume, see i915_gem_restore_fences(). 286 * 287 * This only works for removing the fence register, on acquisition 288 * the caller must hold the rpm wakeref. The fence register must 289 * be cleared before we can use any other fences to ensure that 290 * the new fences do not overlap the elided clears, confusing HW. 291 */ 292 wakeref = intel_runtime_pm_get_if_in_use(uncore->rpm); 293 if (!wakeref) { 294 GEM_BUG_ON(vma); 295 return 0; 296 } 297 298 WRITE_ONCE(fence->vma, vma); 299 fence_write(fence, vma); 300 301 if (vma) { 302 vma->fence = fence; 303 list_move_tail(&fence->link, &ggtt->fence_list); 304 } 305 306 intel_runtime_pm_put(uncore->rpm, wakeref); 307 return 0; 308} 309 310/** 311 * i915_vma_revoke_fence - force-remove fence for a VMA 312 * @vma: vma to map linearly (not through a fence reg) 313 * 314 * This function force-removes any fence from the given object, which is useful 315 * if the kernel wants to do untiled GTT access. 316 * 317 * Returns: 318 * 319 * 0 on success, negative error code on failure. 320 */ 321int i915_vma_revoke_fence(struct i915_vma *vma) 322{ 323 struct i915_fence_reg *fence = vma->fence; 324 325 lockdep_assert_held(&vma->vm->mutex); 326 if (!fence) 327 return 0; 328 329 if (atomic_read(&fence->pin_count)) 330 return -EBUSY; 331 332 return fence_update(fence, NULL); 333} 334 335static struct i915_fence_reg *fence_find(struct i915_ggtt *ggtt) 336{ 337 struct i915_fence_reg *fence; 338 339 list_for_each_entry(fence, &ggtt->fence_list, link) { 340 GEM_BUG_ON(fence->vma && fence->vma->fence != fence); 341 342 if (atomic_read(&fence->pin_count)) 343 continue; 344 345 return fence; 346 } 347 348 /* Wait for completion of pending flips which consume fences */ 349 if (intel_has_pending_fb_unpin(ggtt->vm.i915)) 350 return ERR_PTR(-EAGAIN); 351 352 return ERR_PTR(-EDEADLK); 353} 354 355int __i915_vma_pin_fence(struct i915_vma *vma) 356{ 357 struct i915_ggtt *ggtt = i915_vm_to_ggtt(vma->vm); 358 struct i915_fence_reg *fence; 359 struct i915_vma *set = i915_gem_object_is_tiled(vma->obj) ? vma : NULL; 360 int err; 361 362 lockdep_assert_held(&vma->vm->mutex); 363 364 /* Just update our place in the LRU if our fence is getting reused. */ 365 if (vma->fence) { 366 fence = vma->fence; 367 GEM_BUG_ON(fence->vma != vma); 368 atomic_inc(&fence->pin_count); 369 if (!fence->dirty) { 370 list_move_tail(&fence->link, &ggtt->fence_list); 371 return 0; 372 } 373 } else if (set) { 374 fence = fence_find(ggtt); 375 if (IS_ERR(fence)) 376 return PTR_ERR(fence); 377 378 GEM_BUG_ON(atomic_read(&fence->pin_count)); 379 atomic_inc(&fence->pin_count); 380 } else { 381 return 0; 382 } 383 384 err = fence_update(fence, set); 385 if (err) 386 goto out_unpin; 387 388 GEM_BUG_ON(fence->vma != set); 389 GEM_BUG_ON(vma->fence != (set ? fence : NULL)); 390 391 if (set) 392 return 0; 393 394out_unpin: 395 atomic_dec(&fence->pin_count); 396 return err; 397} 398 399/** 400 * i915_vma_pin_fence - set up fencing for a vma 401 * @vma: vma to map through a fence reg 402 * 403 * When mapping objects through the GTT, userspace wants to be able to write 404 * to them without having to worry about swizzling if the object is tiled. 405 * This function walks the fence regs looking for a free one for @obj, 406 * stealing one if it can't find any. 407 * 408 * It then sets up the reg based on the object's properties: address, pitch 409 * and tiling format. 410 * 411 * For an untiled surface, this removes any existing fence. 412 * 413 * Returns: 414 * 415 * 0 on success, negative error code on failure. 416 */ 417int i915_vma_pin_fence(struct i915_vma *vma) 418{ 419 int err; 420 421 if (!vma->fence && !i915_gem_object_is_tiled(vma->obj)) 422 return 0; 423 424 /* 425 * Note that we revoke fences on runtime suspend. Therefore the user 426 * must keep the device awake whilst using the fence. 427 */ 428 assert_rpm_wakelock_held(vma->vm->gt->uncore->rpm); 429 GEM_BUG_ON(!i915_vma_is_pinned(vma)); 430 GEM_BUG_ON(!i915_vma_is_ggtt(vma)); 431 432 err = mutex_lock_interruptible(&vma->vm->mutex); 433 if (err) 434 return err; 435 436 err = __i915_vma_pin_fence(vma); 437 mutex_unlock(&vma->vm->mutex); 438 439 return err; 440} 441 442/** 443 * i915_reserve_fence - Reserve a fence for vGPU 444 * @ggtt: Global GTT 445 * 446 * This function walks the fence regs looking for a free one and remove 447 * it from the fence_list. It is used to reserve fence for vGPU to use. 448 */ 449struct i915_fence_reg *i915_reserve_fence(struct i915_ggtt *ggtt) 450{ 451 struct i915_fence_reg *fence; 452 int count; 453 int ret; 454 455 lockdep_assert_held(&ggtt->vm.mutex); 456 457 /* Keep at least one fence available for the display engine. */ 458 count = 0; 459 list_for_each_entry(fence, &ggtt->fence_list, link) 460 count += !atomic_read(&fence->pin_count); 461 if (count <= 1) 462 return ERR_PTR(-ENOSPC); 463 464 fence = fence_find(ggtt); 465 if (IS_ERR(fence)) 466 return fence; 467 468 if (fence->vma) { 469 /* Force-remove fence from VMA */ 470 ret = fence_update(fence, NULL); 471 if (ret) 472 return ERR_PTR(ret); 473 } 474 475 list_del(&fence->link); 476 477 return fence; 478} 479 480/** 481 * i915_unreserve_fence - Reclaim a reserved fence 482 * @fence: the fence reg 483 * 484 * This function add a reserved fence register from vGPU to the fence_list. 485 */ 486void i915_unreserve_fence(struct i915_fence_reg *fence) 487{ 488 struct i915_ggtt *ggtt = fence->ggtt; 489 490 lockdep_assert_held(&ggtt->vm.mutex); 491 492 list_add(&fence->link, &ggtt->fence_list); 493} 494 495/** 496 * i915_gem_restore_fences - restore fence state 497 * @ggtt: Global GTT 498 * 499 * Restore the hw fence state to match the software tracking again, to be called 500 * after a gpu reset and on resume. Note that on runtime suspend we only cancel 501 * the fences, to be reacquired by the user later. 502 */ 503void i915_gem_restore_fences(struct i915_ggtt *ggtt) 504{ 505 int i; 506 507 rcu_read_lock(); /* keep obj alive as we dereference */ 508 for (i = 0; i < ggtt->num_fences; i++) { 509 struct i915_fence_reg *reg = &ggtt->fence_regs[i]; 510 struct i915_vma *vma = READ_ONCE(reg->vma); 511 512 GEM_BUG_ON(vma && vma->fence != reg); 513 514 /* 515 * Commit delayed tiling changes if we have an object still 516 * attached to the fence, otherwise just clear the fence. 517 */ 518 if (vma && !i915_gem_object_is_tiled(vma->obj)) 519 vma = NULL; 520 521 fence_write(reg, vma); 522 } 523 rcu_read_unlock(); 524} 525 526/** 527 * DOC: tiling swizzling details 528 * 529 * The idea behind tiling is to increase cache hit rates by rearranging 530 * pixel data so that a group of pixel accesses are in the same cacheline. 531 * Performance improvement from doing this on the back/depth buffer are on 532 * the order of 30%. 533 * 534 * Intel architectures make this somewhat more complicated, though, by 535 * adjustments made to addressing of data when the memory is in interleaved 536 * mode (matched pairs of DIMMS) to improve memory bandwidth. 537 * For interleaved memory, the CPU sends every sequential 64 bytes 538 * to an alternate memory channel so it can get the bandwidth from both. 539 * 540 * The GPU also rearranges its accesses for increased bandwidth to interleaved 541 * memory, and it matches what the CPU does for non-tiled. However, when tiled 542 * it does it a little differently, since one walks addresses not just in the 543 * X direction but also Y. So, along with alternating channels when bit 544 * 6 of the address flips, it also alternates when other bits flip -- Bits 9 545 * (every 512 bytes, an X tile scanline) and 10 (every two X tile scanlines) 546 * are common to both the 915 and 965-class hardware. 547 * 548 * The CPU also sometimes XORs in higher bits as well, to improve 549 * bandwidth doing strided access like we do so frequently in graphics. This 550 * is called "Channel XOR Randomization" in the MCH documentation. The result 551 * is that the CPU is XORing in either bit 11 or bit 17 to bit 6 of its address 552 * decode. 553 * 554 * All of this bit 6 XORing has an effect on our memory management, 555 * as we need to make sure that the 3d driver can correctly address object 556 * contents. 557 * 558 * If we don't have interleaved memory, all tiling is safe and no swizzling is 559 * required. 560 * 561 * When bit 17 is XORed in, we simply refuse to tile at all. Bit 562 * 17 is not just a page offset, so as we page an object out and back in, 563 * individual pages in it will have different bit 17 addresses, resulting in 564 * each 64 bytes being swapped with its neighbor! 565 * 566 * Otherwise, if interleaved, we have to tell the 3d driver what the address 567 * swizzling it needs to do is, since it's writing with the CPU to the pages 568 * (bit 6 and potentially bit 11 XORed in), and the GPU is reading from the 569 * pages (bit 6, 9, and 10 XORed in), resulting in a cumulative bit swizzling 570 * required by the CPU of XORing in bit 6, 9, 10, and potentially 11, in order 571 * to match what the GPU expects. 572 */ 573 574/** 575 * detect_bit_6_swizzle - detect bit 6 swizzling pattern 576 * @ggtt: Global GGTT 577 * 578 * Detects bit 6 swizzling of address lookup between IGD access and CPU 579 * access through main memory. 580 */ 581static void detect_bit_6_swizzle(struct i915_ggtt *ggtt) 582{ 583 struct intel_uncore *uncore = ggtt->vm.gt->uncore; 584 struct drm_i915_private *i915 = ggtt->vm.i915; 585 u32 swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN; 586 u32 swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN; 587 588 if (INTEL_GEN(i915) >= 8 || IS_VALLEYVIEW(i915)) { 589 /* 590 * On BDW+, swizzling is not used. We leave the CPU memory 591 * controller in charge of optimizing memory accesses without 592 * the extra address manipulation GPU side. 593 * 594 * VLV and CHV don't have GPU swizzling. 595 */ 596 swizzle_x = I915_BIT_6_SWIZZLE_NONE; 597 swizzle_y = I915_BIT_6_SWIZZLE_NONE; 598 } else if (INTEL_GEN(i915) >= 6) { 599 if (i915->preserve_bios_swizzle) { 600 if (intel_uncore_read(uncore, DISP_ARB_CTL) & 601 DISP_TILE_SURFACE_SWIZZLING) { 602 swizzle_x = I915_BIT_6_SWIZZLE_9_10; 603 swizzle_y = I915_BIT_6_SWIZZLE_9; 604 } else { 605 swizzle_x = I915_BIT_6_SWIZZLE_NONE; 606 swizzle_y = I915_BIT_6_SWIZZLE_NONE; 607 } 608 } else { 609 u32 dimm_c0, dimm_c1; 610 dimm_c0 = intel_uncore_read(uncore, MAD_DIMM_C0); 611 dimm_c1 = intel_uncore_read(uncore, MAD_DIMM_C1); 612 dimm_c0 &= MAD_DIMM_A_SIZE_MASK | MAD_DIMM_B_SIZE_MASK; 613 dimm_c1 &= MAD_DIMM_A_SIZE_MASK | MAD_DIMM_B_SIZE_MASK; 614 /* 615 * Enable swizzling when the channels are populated 616 * with identically sized dimms. We don't need to check 617 * the 3rd channel because no cpu with gpu attached 618 * ships in that configuration. Also, swizzling only 619 * makes sense for 2 channels anyway. 620 */ 621 if (dimm_c0 == dimm_c1) { 622 swizzle_x = I915_BIT_6_SWIZZLE_9_10; 623 swizzle_y = I915_BIT_6_SWIZZLE_9; 624 } else { 625 swizzle_x = I915_BIT_6_SWIZZLE_NONE; 626 swizzle_y = I915_BIT_6_SWIZZLE_NONE; 627 } 628 } 629 } else if (IS_GEN(i915, 5)) { 630 /* 631 * On Ironlake whatever DRAM config, GPU always do 632 * same swizzling setup. 633 */ 634 swizzle_x = I915_BIT_6_SWIZZLE_9_10; 635 swizzle_y = I915_BIT_6_SWIZZLE_9; 636 } else if (IS_GEN(i915, 2)) { 637 /* 638 * As far as we know, the 865 doesn't have these bit 6 639 * swizzling issues. 640 */ 641 swizzle_x = I915_BIT_6_SWIZZLE_NONE; 642 swizzle_y = I915_BIT_6_SWIZZLE_NONE; 643 } else if (IS_G45(i915) || IS_I965G(i915) || IS_G33(i915)) { 644 /* 645 * The 965, G33, and newer, have a very flexible memory 646 * configuration. It will enable dual-channel mode 647 * (interleaving) on as much memory as it can, and the GPU 648 * will additionally sometimes enable different bit 6 649 * swizzling for tiled objects from the CPU. 650 * 651 * Here's what I found on the G965: 652 * slot fill memory size swizzling 653 * 0A 0B 1A 1B 1-ch 2-ch 654 * 512 0 0 0 512 0 O 655 * 512 0 512 0 16 1008 X 656 * 512 0 0 512 16 1008 X 657 * 0 512 0 512 16 1008 X 658 * 1024 1024 1024 0 2048 1024 O 659 * 660 * We could probably detect this based on either the DRB 661 * matching, which was the case for the swizzling required in 662 * the table above, or from the 1-ch value being less than 663 * the minimum size of a rank. 664 * 665 * Reports indicate that the swizzling actually 666 * varies depending upon page placement inside the 667 * channels, i.e. we see swizzled pages where the 668 * banks of memory are paired and unswizzled on the 669 * uneven portion, so leave that as unknown. 670 */ 671 if (intel_uncore_read(uncore, C0DRB3) == 672 intel_uncore_read(uncore, C1DRB3)) { 673 swizzle_x = I915_BIT_6_SWIZZLE_9_10; 674 swizzle_y = I915_BIT_6_SWIZZLE_9; 675 } 676 } else { 677 u32 dcc = intel_uncore_read(uncore, DCC); 678 679 /* 680 * On 9xx chipsets, channel interleave by the CPU is 681 * determined by DCC. For single-channel, neither the CPU 682 * nor the GPU do swizzling. For dual channel interleaved, 683 * the GPU's interleave is bit 9 and 10 for X tiled, and bit 684 * 9 for Y tiled. The CPU's interleave is independent, and 685 * can be based on either bit 11 (haven't seen this yet) or 686 * bit 17 (common). 687 */ 688 switch (dcc & DCC_ADDRESSING_MODE_MASK) { 689 case DCC_ADDRESSING_MODE_SINGLE_CHANNEL: 690 case DCC_ADDRESSING_MODE_DUAL_CHANNEL_ASYMMETRIC: 691 swizzle_x = I915_BIT_6_SWIZZLE_NONE; 692 swizzle_y = I915_BIT_6_SWIZZLE_NONE; 693 break; 694 case DCC_ADDRESSING_MODE_DUAL_CHANNEL_INTERLEAVED: 695 if (dcc & DCC_CHANNEL_XOR_DISABLE) { 696 /* 697 * This is the base swizzling by the GPU for 698 * tiled buffers. 699 */ 700 swizzle_x = I915_BIT_6_SWIZZLE_9_10; 701 swizzle_y = I915_BIT_6_SWIZZLE_9; 702 } else if ((dcc & DCC_CHANNEL_XOR_BIT_17) == 0) { 703 /* Bit 11 swizzling by the CPU in addition. */ 704 swizzle_x = I915_BIT_6_SWIZZLE_9_10_11; 705 swizzle_y = I915_BIT_6_SWIZZLE_9_11; 706 } else { 707 /* Bit 17 swizzling by the CPU in addition. */ 708 swizzle_x = I915_BIT_6_SWIZZLE_9_10_17; 709 swizzle_y = I915_BIT_6_SWIZZLE_9_17; 710 } 711 break; 712 } 713 714 /* check for L-shaped memory aka modified enhanced addressing */ 715 if (IS_GEN(i915, 4) && 716 !(intel_uncore_read(uncore, DCC2) & DCC2_MODIFIED_ENHANCED_DISABLE)) { 717 swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN; 718 swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN; 719 } 720 721 if (dcc == 0xffffffff) { 722 DRM_ERROR("Couldn't read from MCHBAR. " 723 "Disabling tiling.\n"); 724 swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN; 725 swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN; 726 } 727 } 728 729 if (swizzle_x == I915_BIT_6_SWIZZLE_UNKNOWN || 730 swizzle_y == I915_BIT_6_SWIZZLE_UNKNOWN) { 731 /* 732 * Userspace likes to explode if it sees unknown swizzling, 733 * so lie. We will finish the lie when reporting through 734 * the get-tiling-ioctl by reporting the physical swizzle 735 * mode as unknown instead. 736 * 737 * As we don't strictly know what the swizzling is, it may be 738 * bit17 dependent, and so we need to also prevent the pages 739 * from being moved. 740 */ 741 i915->quirks |= QUIRK_PIN_SWIZZLED_PAGES; 742 swizzle_x = I915_BIT_6_SWIZZLE_NONE; 743 swizzle_y = I915_BIT_6_SWIZZLE_NONE; 744 } 745 746 i915->ggtt.bit_6_swizzle_x = swizzle_x; 747 i915->ggtt.bit_6_swizzle_y = swizzle_y; 748} 749 750/* 751 * Swap every 64 bytes of this page around, to account for it having a new 752 * bit 17 of its physical address and therefore being interpreted differently 753 * by the GPU. 754 */ 755static void i915_gem_swizzle_page(struct page *page) 756{ 757 char temp[64]; 758 char *vaddr; 759 int i; 760 761 vaddr = kmap(page); 762 763 for (i = 0; i < PAGE_SIZE; i += 128) { 764 memcpy(temp, &vaddr[i], 64); 765 memcpy(&vaddr[i], &vaddr[i + 64], 64); 766 memcpy(&vaddr[i + 64], temp, 64); 767 } 768 769 kunmap(page); 770} 771 772/** 773 * i915_gem_object_do_bit_17_swizzle - fixup bit 17 swizzling 774 * @obj: i915 GEM buffer object 775 * @pages: the scattergather list of physical pages 776 * 777 * This function fixes up the swizzling in case any page frame number for this 778 * object has changed in bit 17 since that state has been saved with 779 * i915_gem_object_save_bit_17_swizzle(). 780 * 781 * This is called when pinning backing storage again, since the kernel is free 782 * to move unpinned backing storage around (either by directly moving pages or 783 * by swapping them out and back in again). 784 */ 785void 786i915_gem_object_do_bit_17_swizzle(struct drm_i915_gem_object *obj, 787 struct sg_table *pages) 788{ 789 struct sgt_iter sgt_iter; 790 struct page *page; 791 int i; 792 793 if (obj->bit_17 == NULL) 794 return; 795 796 i = 0; 797 for_each_sgt_page(page, sgt_iter, pages) { 798 char new_bit_17 = page_to_phys(page) >> 17; 799 if ((new_bit_17 & 0x1) != (test_bit(i, obj->bit_17) != 0)) { 800 i915_gem_swizzle_page(page); 801 set_page_dirty(page); 802 } 803 i++; 804 } 805} 806 807/** 808 * i915_gem_object_save_bit_17_swizzle - save bit 17 swizzling 809 * @obj: i915 GEM buffer object 810 * @pages: the scattergather list of physical pages 811 * 812 * This function saves the bit 17 of each page frame number so that swizzling 813 * can be fixed up later on with i915_gem_object_do_bit_17_swizzle(). This must 814 * be called before the backing storage can be unpinned. 815 */ 816void 817i915_gem_object_save_bit_17_swizzle(struct drm_i915_gem_object *obj, 818 struct sg_table *pages) 819{ 820 const unsigned int page_count = obj->base.size >> PAGE_SHIFT; 821 struct sgt_iter sgt_iter; 822 struct page *page; 823 int i; 824 825 if (obj->bit_17 == NULL) { 826 obj->bit_17 = bitmap_zalloc(page_count, GFP_KERNEL); 827 if (obj->bit_17 == NULL) { 828 DRM_ERROR("Failed to allocate memory for bit 17 " 829 "record\n"); 830 return; 831 } 832 } 833 834 i = 0; 835 836 for_each_sgt_page(page, sgt_iter, pages) { 837 if (page_to_phys(page) & (1 << 17)) 838 __set_bit(i, obj->bit_17); 839 else 840 __clear_bit(i, obj->bit_17); 841 i++; 842 } 843} 844 845void i915_ggtt_init_fences(struct i915_ggtt *ggtt) 846{ 847 struct drm_i915_private *i915 = ggtt->vm.i915; 848 struct intel_uncore *uncore = ggtt->vm.gt->uncore; 849 int num_fences; 850 int i; 851 852 INIT_LIST_HEAD(&ggtt->fence_list); 853 INIT_LIST_HEAD(&ggtt->userfault_list); 854 intel_wakeref_auto_init(&ggtt->userfault_wakeref, uncore->rpm); 855 856 detect_bit_6_swizzle(ggtt); 857 858 if (!i915_ggtt_has_aperture(ggtt)) 859 num_fences = 0; 860 else if (INTEL_GEN(i915) >= 7 && 861 !(IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915))) 862 num_fences = 32; 863 else if (INTEL_GEN(i915) >= 4 || 864 IS_I945G(i915) || IS_I945GM(i915) || 865 IS_G33(i915) || IS_PINEVIEW(i915)) 866 num_fences = 16; 867 else 868 num_fences = 8; 869 870 if (intel_vgpu_active(i915)) 871 num_fences = intel_uncore_read(uncore, 872 vgtif_reg(avail_rs.fence_num)); 873 874 /* Initialize fence registers to zero */ 875 for (i = 0; i < num_fences; i++) { 876 struct i915_fence_reg *fence = &ggtt->fence_regs[i]; 877 878 fence->ggtt = ggtt; 879 fence->id = i; 880 list_add_tail(&fence->link, &ggtt->fence_list); 881 } 882 ggtt->num_fences = num_fences; 883 884 i915_gem_restore_fences(ggtt); 885} 886 887void intel_gt_init_swizzling(struct intel_gt *gt) 888{ 889 struct drm_i915_private *i915 = gt->i915; 890 struct intel_uncore *uncore = gt->uncore; 891 892 if (INTEL_GEN(i915) < 5 || 893 i915->ggtt.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE) 894 return; 895 896 intel_uncore_rmw(uncore, DISP_ARB_CTL, 0, DISP_TILE_SURFACE_SWIZZLING); 897 898 if (IS_GEN(i915, 5)) 899 return; 900 901 intel_uncore_rmw(uncore, TILECTL, 0, TILECTL_SWZCTL); 902 903 if (IS_GEN(i915, 6)) 904 intel_uncore_write(uncore, 905 ARB_MODE, 906 _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB)); 907 else if (IS_GEN(i915, 7)) 908 intel_uncore_write(uncore, 909 ARB_MODE, 910 _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB)); 911 else if (IS_GEN(i915, 8)) 912 intel_uncore_write(uncore, 913 GAMTARBMODE, 914 _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW)); 915 else 916 MISSING_CASE(INTEL_GEN(i915)); 917} 918