projects / linux-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
drm/i915/bdw: Initialize the PDEs
[linux-block.git] / drivers / gpu / drm / i915 / i915_gem_gtt.c
CommitLineData
76aaf220
DV		 1/*
2 * Copyright © 2010 Daniel Vetter
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 *
23 */
24
760285e7
DH		 25#include <drm/drmP.h>
26#include <drm/i915_drm.h>
76aaf220
DV		 27#include "i915_drv.h"
28#include "i915_trace.h"
29#include "intel_drv.h"
30
6670a5a5
BW		 31#define GEN6_PPGTT_PD_ENTRIES 512
32#define I915_PPGTT_PT_ENTRIES (PAGE_SIZE / sizeof(gen6_gtt_pte_t))
d31eb10e 33typedef uint64_t gen8_gtt_pte_t;
37aca44a 34typedef gen8_gtt_pte_t gen8_ppgtt_pde_t;
6670a5a5 35
26b1ff35
BW		 36/* PPGTT stuff */
37#define GEN6_GTT_ADDR_ENCODE(addr) ((addr) | (((addr) >> 28) & 0xff0))
0d8ff15e 38#define HSW_GTT_ADDR_ENCODE(addr) ((addr) | (((addr) >> 28) & 0x7f0))
26b1ff35
BW		 39
40#define GEN6_PDE_VALID (1 << 0)
41/* gen6+ has bit 11-4 for physical addr bit 39-32 */
42#define GEN6_PDE_ADDR_ENCODE(addr) GEN6_GTT_ADDR_ENCODE(addr)
43
44#define GEN6_PTE_VALID (1 << 0)
45#define GEN6_PTE_UNCACHED (1 << 1)
46#define HSW_PTE_UNCACHED (0)
47#define GEN6_PTE_CACHE_LLC (2 << 1)
350ec881 48#define GEN7_PTE_CACHE_L3_LLC (3 << 1)
26b1ff35 49#define GEN6_PTE_ADDR_ENCODE(addr) GEN6_GTT_ADDR_ENCODE(addr)
0d8ff15e
BW		 50#define HSW_PTE_ADDR_ENCODE(addr) HSW_GTT_ADDR_ENCODE(addr)
51
52/* Cacheability Control is a 4-bit value. The low three bits are stored in *
53 * bits 3:1 of the PTE, while the fourth bit is stored in bit 11 of the PTE.
54 */
55#define HSW_CACHEABILITY_CONTROL(bits) ((((bits) & 0x7) << 1) | \
56 (((bits) & 0x8) << (11 - 3)))
87a6b688 57#define HSW_WB_LLC_AGE3 HSW_CACHEABILITY_CONTROL(0x2)
0d8ff15e 58#define HSW_WB_LLC_AGE0 HSW_CACHEABILITY_CONTROL(0x3)
4d15c145 59#define HSW_WB_ELLC_LLC_AGE0 HSW_CACHEABILITY_CONTROL(0xb)
651d794f 60#define HSW_WT_ELLC_LLC_AGE0 HSW_CACHEABILITY_CONTROL(0x6)
26b1ff35 61
37aca44a
BW		 62#define GEN8_PDES_PER_PAGE (PAGE_SIZE / sizeof(gen8_ppgtt_pde_t))
63#define GEN8_LEGACY_PDPS 4
64
fbe5d36e
BW		 65#define PPAT_UNCACHED_INDEX (_PAGE_PWT | _PAGE_PCD)
66#define PPAT_CACHED_PDE_INDEX 0 /* WB LLC */
67#define PPAT_CACHED_INDEX _PAGE_PAT /* WB LLCeLLC */
68#define PPAT_DISPLAY_ELLC_INDEX _PAGE_PCD /* WT eLLC */
69
94ec8f61
BW		 70static inline gen8_gtt_pte_t gen8_pte_encode(dma_addr_t addr,
71 enum i915_cache_level level,
72 bool valid)
73{
74 gen8_gtt_pte_t pte = valid ? _PAGE_PRESENT | _PAGE_RW : 0;
75 pte |= addr;
fbe5d36e
BW		 76 if (level != I915_CACHE_NONE)
77 pte |= PPAT_CACHED_INDEX;
78 else
79 pte |= PPAT_UNCACHED_INDEX;
94ec8f61
BW		 80 return pte;
81}
82
b1fe6673
BW		 83static inline gen8_ppgtt_pde_t gen8_pde_encode(struct drm_device *dev,
84 dma_addr_t addr,
85 enum i915_cache_level level)
86{
87 gen8_ppgtt_pde_t pde = _PAGE_PRESENT | _PAGE_RW;
88 pde |= addr;
89 if (level != I915_CACHE_NONE)
90 pde |= PPAT_CACHED_PDE_INDEX;
91 else
92 pde |= PPAT_UNCACHED_INDEX;
93 return pde;
94}
95
350ec881 96static gen6_gtt_pte_t snb_pte_encode(dma_addr_t addr,
b35b380e
BW		 97 enum i915_cache_level level,
98 bool valid)
54d12527 99{
b35b380e 100 gen6_gtt_pte_t pte = valid ? GEN6_PTE_VALID : 0;
54d12527 101 pte |= GEN6_PTE_ADDR_ENCODE(addr);
e7210c3c
BW		 102
103 switch (level) {
350ec881
CW		 104 case I915_CACHE_L3_LLC:
105 case I915_CACHE_LLC:
106 pte |= GEN6_PTE_CACHE_LLC;
107 break;
108 case I915_CACHE_NONE:
109 pte |= GEN6_PTE_UNCACHED;
110 break;
111 default:
112 WARN_ON(1);
113 }
114
115 return pte;
116}
117
118static gen6_gtt_pte_t ivb_pte_encode(dma_addr_t addr,
b35b380e
BW		 119 enum i915_cache_level level,
120 bool valid)
350ec881 121{
b35b380e 122 gen6_gtt_pte_t pte = valid ? GEN6_PTE_VALID : 0;
350ec881
CW		 123 pte |= GEN6_PTE_ADDR_ENCODE(addr);
124
125 switch (level) {
126 case I915_CACHE_L3_LLC:
127 pte |= GEN7_PTE_CACHE_L3_LLC;
e7210c3c
BW		 128 break;
129 case I915_CACHE_LLC:
130 pte |= GEN6_PTE_CACHE_LLC;
131 break;
132 case I915_CACHE_NONE:
9119708c 133 pte |= GEN6_PTE_UNCACHED;
e7210c3c
BW		 134 break;
135 default:
350ec881 136 WARN_ON(1);
e7210c3c
BW		 137 }
138
54d12527
BW		 139 return pte;
140}
141
93c34e70
KG		 142#define BYT_PTE_WRITEABLE (1 << 1)
143#define BYT_PTE_SNOOPED_BY_CPU_CACHES (1 << 2)
144
80a74f7f 145static gen6_gtt_pte_t byt_pte_encode(dma_addr_t addr,
b35b380e
BW		 146 enum i915_cache_level level,
147 bool valid)
93c34e70 148{
b35b380e 149 gen6_gtt_pte_t pte = valid ? GEN6_PTE_VALID : 0;
93c34e70
KG		 150 pte |= GEN6_PTE_ADDR_ENCODE(addr);
151
152 /* Mark the page as writeable. Other platforms don't have a
153 * setting for read-only/writable, so this matches that behavior.
154 */
155 pte |= BYT_PTE_WRITEABLE;
156
157 if (level != I915_CACHE_NONE)
158 pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
159
160 return pte;
161}
162
80a74f7f 163static gen6_gtt_pte_t hsw_pte_encode(dma_addr_t addr,
b35b380e
BW		 164 enum i915_cache_level level,
165 bool valid)
9119708c 166{
b35b380e 167 gen6_gtt_pte_t pte = valid ? GEN6_PTE_VALID : 0;
0d8ff15e 168 pte |= HSW_PTE_ADDR_ENCODE(addr);
9119708c
KG		 169
170 if (level != I915_CACHE_NONE)
87a6b688 171 pte |= HSW_WB_LLC_AGE3;
9119708c
KG		 172
173 return pte;
174}
175
4d15c145 176static gen6_gtt_pte_t iris_pte_encode(dma_addr_t addr,
b35b380e
BW		 177 enum i915_cache_level level,
178 bool valid)
4d15c145 179{
b35b380e 180 gen6_gtt_pte_t pte = valid ? GEN6_PTE_VALID : 0;
4d15c145
BW		 181 pte |= HSW_PTE_ADDR_ENCODE(addr);
182
651d794f
CW		 183 switch (level) {
184 case I915_CACHE_NONE:
185 break;
186 case I915_CACHE_WT:
187 pte |= HSW_WT_ELLC_LLC_AGE0;
188 break;
189 default:
4d15c145 190 pte |= HSW_WB_ELLC_LLC_AGE0;
651d794f
CW		 191 break;
192 }
4d15c145
BW		 193
194 return pte;
195}
196
37aca44a
BW		 197static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
198{
199 struct i915_hw_ppgtt *ppgtt =
200 container_of(vm, struct i915_hw_ppgtt, base);
201 int i, j;
202
203 for (i = 0; i < ppgtt->num_pd_pages ; i++) {
204 if (ppgtt->pd_dma_addr[i]) {
205 pci_unmap_page(ppgtt->base.dev->pdev,
206 ppgtt->pd_dma_addr[i],
207 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
208
209 for (j = 0; j < GEN8_PDES_PER_PAGE; j++) {
210 dma_addr_t addr = ppgtt->gen8_pt_dma_addr[i][j];
211 if (addr)
212 pci_unmap_page(ppgtt->base.dev->pdev,
213 addr,
214 PAGE_SIZE,
215 PCI_DMA_BIDIRECTIONAL);
216
217 }
218 }
219 kfree(ppgtt->gen8_pt_dma_addr[i]);
220 }
221
222 __free_pages(ppgtt->gen8_pt_pages, ppgtt->num_pt_pages << PAGE_SHIFT);
223 __free_pages(ppgtt->pd_pages, ppgtt->num_pd_pages << PAGE_SHIFT);
224}
225
226/**
227 * GEN8 legacy ppgtt programming is accomplished through 4 PDP registers with a
228 * net effect resembling a 2-level page table in normal x86 terms. Each PDP
229 * represents 1GB of memory
230 * 4 * 512 * 512 * 4096 = 4GB legacy 32b address space.
231 *
232 * TODO: Do something with the size parameter
233 **/
234static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt, uint64_t size)
235{
236 struct page *pt_pages;
237 int i, j, ret = -ENOMEM;
238 const int max_pdp = DIV_ROUND_UP(size, 1 << 30);
239 const int num_pt_pages = GEN8_PDES_PER_PAGE * max_pdp;
240
241 if (size % (1<<30))
242 DRM_INFO("Pages will be wasted unless GTT size (%llu) is divisible by 1GB\n", size);
243
244 /* FIXME: split allocation into smaller pieces. For now we only ever do
245 * this once, but with full PPGTT, the multiple contiguous allocations
246 * will be bad.
247 */
248 ppgtt->pd_pages = alloc_pages(GFP_KERNEL, get_order(max_pdp << PAGE_SHIFT));
249 if (!ppgtt->pd_pages)
250 return -ENOMEM;
251
252 pt_pages = alloc_pages(GFP_KERNEL, get_order(num_pt_pages << PAGE_SHIFT));
253 if (!pt_pages) {
254 __free_pages(ppgtt->pd_pages, get_order(max_pdp << PAGE_SHIFT));
255 return -ENOMEM;
256 }
257
258 ppgtt->gen8_pt_pages = pt_pages;
259 ppgtt->num_pd_pages = 1 << get_order(max_pdp << PAGE_SHIFT);
260 ppgtt->num_pt_pages = 1 << get_order(num_pt_pages << PAGE_SHIFT);
261 ppgtt->num_pd_entries = max_pdp * GEN8_PDES_PER_PAGE;
262 ppgtt->base.clear_range = NULL;
263 ppgtt->base.insert_entries = NULL;
264 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
265
266 BUG_ON(ppgtt->num_pd_pages > GEN8_LEGACY_PDPS);
267
268 /*
269 * - Create a mapping for the page directories.
270 * - For each page directory:
271 * allocate space for page table mappings.
272 * map each page table
273 */
274 for (i = 0; i < max_pdp; i++) {
275 dma_addr_t temp;
276 temp = pci_map_page(ppgtt->base.dev->pdev,
277 &ppgtt->pd_pages[i], 0,
278 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
279 if (pci_dma_mapping_error(ppgtt->base.dev->pdev, temp))
280 goto err_out;
281
282 ppgtt->pd_dma_addr[i] = temp;
283
284 ppgtt->gen8_pt_dma_addr[i] = kmalloc(sizeof(dma_addr_t) * GEN8_PDES_PER_PAGE, GFP_KERNEL);
285 if (!ppgtt->gen8_pt_dma_addr[i])
286 goto err_out;
287
288 for (j = 0; j < GEN8_PDES_PER_PAGE; j++) {
289 struct page *p = &pt_pages[i * GEN8_PDES_PER_PAGE + j];
290 temp = pci_map_page(ppgtt->base.dev->pdev,
291 p, 0, PAGE_SIZE,
292 PCI_DMA_BIDIRECTIONAL);
293
294 if (pci_dma_mapping_error(ppgtt->base.dev->pdev, temp))
295 goto err_out;
296
297 ppgtt->gen8_pt_dma_addr[i][j] = temp;
298 }
299 }
300
b1fe6673
BW		 301 /* For now, the PPGTT helper functions all require that the PDEs are
302 * plugged in correctly. So we do that now/here. For aliasing PPGTT, we
303 * will never need to touch the PDEs again */
304 for (i = 0; i < max_pdp; i++) {
305 gen8_ppgtt_pde_t *pd_vaddr;
306 pd_vaddr = kmap_atomic(&ppgtt->pd_pages[i]);
307 for (j = 0; j < GEN8_PDES_PER_PAGE; j++) {
308 dma_addr_t addr = ppgtt->gen8_pt_dma_addr[i][j];
309 pd_vaddr[j] = gen8_pde_encode(ppgtt->base.dev, addr,
310 I915_CACHE_LLC);
311 }
312 kunmap_atomic(pd_vaddr);
313 }
314
37aca44a
BW		 315 DRM_DEBUG_DRIVER("Allocated %d pages for page directories (%d wasted)\n",
316 ppgtt->num_pd_pages, ppgtt->num_pd_pages - max_pdp);
317 DRM_DEBUG_DRIVER("Allocated %d pages for page tables (%lld wasted)\n",
318 ppgtt->num_pt_pages,
319 (ppgtt->num_pt_pages - num_pt_pages) +
320 size % (1<<30));
321 return -ENOSYS; /* Not ready yet */
322
323err_out:
324 ppgtt->base.cleanup(&ppgtt->base);
325 return ret;
326}
327
3e302542 328static void gen6_write_pdes(struct i915_hw_ppgtt *ppgtt)
6197349b 329{
853ba5d2 330 struct drm_i915_private *dev_priv = ppgtt->base.dev->dev_private;
6197349b
BW		 331 gen6_gtt_pte_t __iomem *pd_addr;
332 uint32_t pd_entry;
333 int i;
334
0a732870 335 WARN_ON(ppgtt->pd_offset & 0x3f);
6197349b
BW		 336 pd_addr = (gen6_gtt_pte_t __iomem*)dev_priv->gtt.gsm +
337 ppgtt->pd_offset / sizeof(gen6_gtt_pte_t);
338 for (i = 0; i < ppgtt->num_pd_entries; i++) {
339 dma_addr_t pt_addr;
340
341 pt_addr = ppgtt->pt_dma_addr[i];
342 pd_entry = GEN6_PDE_ADDR_ENCODE(pt_addr);
343 pd_entry |= GEN6_PDE_VALID;
344
345 writel(pd_entry, pd_addr + i);
346 }
347 readl(pd_addr);
3e302542
BW		 348}
349
350static int gen6_ppgtt_enable(struct drm_device *dev)
351{
352 drm_i915_private_t *dev_priv = dev->dev_private;
353 uint32_t pd_offset;
354 struct intel_ring_buffer *ring;
355 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
356 int i;
357
358 BUG_ON(ppgtt->pd_offset & 0x3f);
359
360 gen6_write_pdes(ppgtt);
6197349b
BW		 361
362 pd_offset = ppgtt->pd_offset;
363 pd_offset /= 64; /* in cachelines, */
364 pd_offset <<= 16;
365
366 if (INTEL_INFO(dev)->gen == 6) {
367 uint32_t ecochk, gab_ctl, ecobits;
368
369 ecobits = I915_READ(GAC_ECO_BITS);
3b9d7888
VS		 370 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
371 ECOBITS_PPGTT_CACHE64B);
6197349b
BW		 372
373 gab_ctl = I915_READ(GAB_CTL);
374 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
375
376 ecochk = I915_READ(GAM_ECOCHK);
377 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT |
378 ECOCHK_PPGTT_CACHE64B);
379 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
380 } else if (INTEL_INFO(dev)->gen >= 7) {
a6f429a5 381 uint32_t ecochk, ecobits;
a65c2fcd
VS		 382
383 ecobits = I915_READ(GAC_ECO_BITS);
384 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
385
a6f429a5
VS		 386 ecochk = I915_READ(GAM_ECOCHK);
387 if (IS_HASWELL(dev)) {
388 ecochk |= ECOCHK_PPGTT_WB_HSW;
389 } else {
390 ecochk |= ECOCHK_PPGTT_LLC_IVB;
391 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
392 }
393 I915_WRITE(GAM_ECOCHK, ecochk);
6197349b
BW		 394 /* GFX_MODE is per-ring on gen7+ */
395 }
396
397 for_each_ring(ring, dev_priv, i) {
398 if (INTEL_INFO(dev)->gen >= 7)
399 I915_WRITE(RING_MODE_GEN7(ring),
400 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
401
402 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
403 I915_WRITE(RING_PP_DIR_BASE(ring), pd_offset);
404 }
b7c36d25 405 return 0;
6197349b
BW		 406}
407
1d2a314c 408/* PPGTT support for Sandybdrige/Gen6 and later */
853ba5d2 409static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1d2a314c 410 unsigned first_entry,
828c7908
BW		 411 unsigned num_entries,
412 bool use_scratch)
1d2a314c 413{
853ba5d2
BW		 414 struct i915_hw_ppgtt *ppgtt =
415 container_of(vm, struct i915_hw_ppgtt, base);
e7c2b58b 416 gen6_gtt_pte_t *pt_vaddr, scratch_pte;
a15326a5 417 unsigned act_pt = first_entry / I915_PPGTT_PT_ENTRIES;
7bddb01f
DV		 418 unsigned first_pte = first_entry % I915_PPGTT_PT_ENTRIES;
419 unsigned last_pte, i;
1d2a314c 420
b35b380e 421 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, true);
1d2a314c 422
7bddb01f
DV		 423 while (num_entries) {
424 last_pte = first_pte + num_entries;
425 if (last_pte > I915_PPGTT_PT_ENTRIES)
426 last_pte = I915_PPGTT_PT_ENTRIES;
427
a15326a5 428 pt_vaddr = kmap_atomic(ppgtt->pt_pages[act_pt]);
1d2a314c 429
7bddb01f
DV		 430 for (i = first_pte; i < last_pte; i++)
431 pt_vaddr[i] = scratch_pte;
1d2a314c
DV		 432
433 kunmap_atomic(pt_vaddr);
1d2a314c 434
7bddb01f
DV		 435 num_entries -= last_pte - first_pte;
436 first_pte = 0;
a15326a5 437 act_pt++;
7bddb01f 438 }
1d2a314c
DV		 439}
440
853ba5d2 441static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
def886c3
DV		 442 struct sg_table *pages,
443 unsigned first_entry,
444 enum i915_cache_level cache_level)
445{
853ba5d2
BW		 446 struct i915_hw_ppgtt *ppgtt =
447 container_of(vm, struct i915_hw_ppgtt, base);
e7c2b58b 448 gen6_gtt_pte_t *pt_vaddr;
a15326a5 449 unsigned act_pt = first_entry / I915_PPGTT_PT_ENTRIES;
6e995e23
ID		 450 unsigned act_pte = first_entry % I915_PPGTT_PT_ENTRIES;
451 struct sg_page_iter sg_iter;
452
a15326a5 453 pt_vaddr = kmap_atomic(ppgtt->pt_pages[act_pt]);
6e995e23
ID		 454 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
455 dma_addr_t page_addr;
456
2db76d7c 457 page_addr = sg_page_iter_dma_address(&sg_iter);
b35b380e 458 pt_vaddr[act_pte] = vm->pte_encode(page_addr, cache_level, true);
6e995e23
ID		 459 if (++act_pte == I915_PPGTT_PT_ENTRIES) {
460 kunmap_atomic(pt_vaddr);
a15326a5
DV		 461 act_pt++;
462 pt_vaddr = kmap_atomic(ppgtt->pt_pages[act_pt]);
6e995e23 463 act_pte = 0;
def886c3 464
def886c3 465 }
def886c3 466 }
6e995e23 467 kunmap_atomic(pt_vaddr);
def886c3
DV		 468}
469
853ba5d2 470static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
1d2a314c 471{
853ba5d2
BW		 472 struct i915_hw_ppgtt *ppgtt =
473 container_of(vm, struct i915_hw_ppgtt, base);
3440d265
DV		 474 int i;
475
93bd8649
BW		 476 drm_mm_takedown(&ppgtt->base.mm);
477
3440d265
DV		 478 if (ppgtt->pt_dma_addr) {
479 for (i = 0; i < ppgtt->num_pd_entries; i++)
853ba5d2 480 pci_unmap_page(ppgtt->base.dev->pdev,
3440d265
DV		 481 ppgtt->pt_dma_addr[i],
482 4096, PCI_DMA_BIDIRECTIONAL);
483 }
484
485 kfree(ppgtt->pt_dma_addr);
486 for (i = 0; i < ppgtt->num_pd_entries; i++)
487 __free_page(ppgtt->pt_pages[i]);
488 kfree(ppgtt->pt_pages);
489 kfree(ppgtt);
490}
491
492static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
493{
853ba5d2 494 struct drm_device *dev = ppgtt->base.dev;
1d2a314c 495 struct drm_i915_private *dev_priv = dev->dev_private;
1d2a314c 496 unsigned first_pd_entry_in_global_pt;
1d2a314c
DV		 497 int i;
498 int ret = -ENOMEM;
499
500 /* ppgtt PDEs reside in the global gtt pagetable, which has 512*1024
501 * entries. For aliasing ppgtt support we just steal them at the end for
502 * now. */
e1b73cba 503 first_pd_entry_in_global_pt = gtt_total_entries(dev_priv->gtt);
1d2a314c 504
08c45263 505 ppgtt->base.pte_encode = dev_priv->gtt.base.pte_encode;
6670a5a5 506 ppgtt->num_pd_entries = GEN6_PPGTT_PD_ENTRIES;
6197349b 507 ppgtt->enable = gen6_ppgtt_enable;
853ba5d2
BW		 508 ppgtt->base.clear_range = gen6_ppgtt_clear_range;
509 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
510 ppgtt->base.cleanup = gen6_ppgtt_cleanup;
511 ppgtt->base.scratch = dev_priv->gtt.base.scratch;
a1e22653 512 ppgtt->pt_pages = kcalloc(ppgtt->num_pd_entries, sizeof(struct page *),
1d2a314c
DV		 513 GFP_KERNEL);
514 if (!ppgtt->pt_pages)
3440d265 515 return -ENOMEM;
1d2a314c
DV		 516
517 for (i = 0; i < ppgtt->num_pd_entries; i++) {
518 ppgtt->pt_pages[i] = alloc_page(GFP_KERNEL);
519 if (!ppgtt->pt_pages[i])
520 goto err_pt_alloc;
521 }
522
a1e22653 523 ppgtt->pt_dma_addr = kcalloc(ppgtt->num_pd_entries, sizeof(dma_addr_t),
8d2e6308
BW		 524 GFP_KERNEL);
525 if (!ppgtt->pt_dma_addr)
526 goto err_pt_alloc;
1d2a314c 527
8d2e6308
BW		 528 for (i = 0; i < ppgtt->num_pd_entries; i++) {
529 dma_addr_t pt_addr;
211c568b 530
8d2e6308
BW		 531 pt_addr = pci_map_page(dev->pdev, ppgtt->pt_pages[i], 0, 4096,
532 PCI_DMA_BIDIRECTIONAL);
1d2a314c 533
8d2e6308
BW		 534 if (pci_dma_mapping_error(dev->pdev, pt_addr)) {
535 ret = -EIO;
536 goto err_pd_pin;
1d2a314c 537
211c568b 538 }
8d2e6308 539 ppgtt->pt_dma_addr[i] = pt_addr;
1d2a314c 540 }
1d2a314c 541
853ba5d2 542 ppgtt->base.clear_range(&ppgtt->base, 0,
828c7908 543 ppgtt->num_pd_entries * I915_PPGTT_PT_ENTRIES, true);
1d2a314c 544
e7c2b58b 545 ppgtt->pd_offset = first_pd_entry_in_global_pt * sizeof(gen6_gtt_pte_t);
1d2a314c 546
1d2a314c
DV		 547 return 0;
548
549err_pd_pin:
550 if (ppgtt->pt_dma_addr) {
551 for (i--; i >= 0; i--)
552 pci_unmap_page(dev->pdev, ppgtt->pt_dma_addr[i],
553 4096, PCI_DMA_BIDIRECTIONAL);
554 }
555err_pt_alloc:
556 kfree(ppgtt->pt_dma_addr);
557 for (i = 0; i < ppgtt->num_pd_entries; i++) {
558 if (ppgtt->pt_pages[i])
559 __free_page(ppgtt->pt_pages[i]);
560 }
561 kfree(ppgtt->pt_pages);
3440d265
DV		 562
563 return ret;
564}
565
566static int i915_gem_init_aliasing_ppgtt(struct drm_device *dev)
567{
568 struct drm_i915_private *dev_priv = dev->dev_private;
569 struct i915_hw_ppgtt *ppgtt;
570 int ret;
571
572 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
573 if (!ppgtt)
574 return -ENOMEM;
575
853ba5d2 576 ppgtt->base.dev = dev;
3440d265 577
3ed124b2
BW		 578 if (INTEL_INFO(dev)->gen < 8)
579 ret = gen6_ppgtt_init(ppgtt);
8fe6bd23 580 else if (IS_GEN8(dev))
37aca44a 581 ret = gen8_ppgtt_init(ppgtt, dev_priv->gtt.base.total);
3ed124b2
BW		 582 else
583 BUG();
584
3440d265
DV		 585 if (ret)
586 kfree(ppgtt);
93bd8649 587 else {
3440d265 588 dev_priv->mm.aliasing_ppgtt = ppgtt;
93bd8649
BW		 589 drm_mm_init(&ppgtt->base.mm, ppgtt->base.start,
590 ppgtt->base.total);
591 }
1d2a314c
DV		 592
593 return ret;
594}
595
596void i915_gem_cleanup_aliasing_ppgtt(struct drm_device *dev)
597{
598 struct drm_i915_private *dev_priv = dev->dev_private;
599 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
1d2a314c
DV		 600
601 if (!ppgtt)
602 return;
603
853ba5d2 604 ppgtt->base.cleanup(&ppgtt->base);
5963cf04 605 dev_priv->mm.aliasing_ppgtt = NULL;
1d2a314c
DV		 606}
607
7bddb01f
DV		 608void i915_ppgtt_bind_object(struct i915_hw_ppgtt *ppgtt,
609 struct drm_i915_gem_object *obj,
610 enum i915_cache_level cache_level)
611{
853ba5d2
BW		 612 ppgtt->base.insert_entries(&ppgtt->base, obj->pages,
613 i915_gem_obj_ggtt_offset(obj) >> PAGE_SHIFT,
614 cache_level);
7bddb01f
DV		 615}
616
617void i915_ppgtt_unbind_object(struct i915_hw_ppgtt *ppgtt,
618 struct drm_i915_gem_object *obj)
619{
853ba5d2
BW		 620 ppgtt->base.clear_range(&ppgtt->base,
621 i915_gem_obj_ggtt_offset(obj) >> PAGE_SHIFT,
828c7908
BW		 622 obj->base.size >> PAGE_SHIFT,
623 true);
7bddb01f
DV		 624}
625
a81cc00c
BW		 626extern int intel_iommu_gfx_mapped;
627/* Certain Gen5 chipsets require require idling the GPU before
628 * unmapping anything from the GTT when VT-d is enabled.
629 */
630static inline bool needs_idle_maps(struct drm_device *dev)
631{
632#ifdef CONFIG_INTEL_IOMMU
633 /* Query intel_iommu to see if we need the workaround. Presumably that
634 * was loaded first.
635 */
636 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
637 return true;
638#endif
639 return false;
640}
641
5c042287
BW		 642static bool do_idling(struct drm_i915_private *dev_priv)
643{
644 bool ret = dev_priv->mm.interruptible;
645
a81cc00c 646 if (unlikely(dev_priv->gtt.do_idle_maps)) {
5c042287 647 dev_priv->mm.interruptible = false;
b2da9fe5 648 if (i915_gpu_idle(dev_priv->dev)) {
5c042287
BW		 649 DRM_ERROR("Couldn't idle GPU\n");
650 /* Wait a bit, in hopes it avoids the hang */
651 udelay(10);
652 }
653 }
654
655 return ret;
656}
657
658static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
659{
a81cc00c 660 if (unlikely(dev_priv->gtt.do_idle_maps))
5c042287
BW		 661 dev_priv->mm.interruptible = interruptible;
662}
663
828c7908
BW		 664void i915_check_and_clear_faults(struct drm_device *dev)
665{
666 struct drm_i915_private *dev_priv = dev->dev_private;
667 struct intel_ring_buffer *ring;
668 int i;
669
670 if (INTEL_INFO(dev)->gen < 6)
671 return;
672
673 for_each_ring(ring, dev_priv, i) {
674 u32 fault_reg;
675 fault_reg = I915_READ(RING_FAULT_REG(ring));
676 if (fault_reg & RING_FAULT_VALID) {
677 DRM_DEBUG_DRIVER("Unexpected fault\n"
678 "\tAddr: 0x%08lx\\n"
679 "\tAddress space: %s\n"
680 "\tSource ID: %d\n"
681 "\tType: %d\n",
682 fault_reg & PAGE_MASK,
683 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
684 RING_FAULT_SRCID(fault_reg),
685 RING_FAULT_FAULT_TYPE(fault_reg));
686 I915_WRITE(RING_FAULT_REG(ring),
687 fault_reg & ~RING_FAULT_VALID);
688 }
689 }
690 POSTING_READ(RING_FAULT_REG(&dev_priv->ring[RCS]));
691}
692
693void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
694{
695 struct drm_i915_private *dev_priv = dev->dev_private;
696
697 /* Don't bother messing with faults pre GEN6 as we have little
698 * documentation supporting that it's a good idea.
699 */
700 if (INTEL_INFO(dev)->gen < 6)
701 return;
702
703 i915_check_and_clear_faults(dev);
704
705 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
706 dev_priv->gtt.base.start / PAGE_SIZE,
707 dev_priv->gtt.base.total / PAGE_SIZE,
708 false);
709}
710
76aaf220
DV		 711void i915_gem_restore_gtt_mappings(struct drm_device *dev)
712{
713 struct drm_i915_private *dev_priv = dev->dev_private;
05394f39 714 struct drm_i915_gem_object *obj;
76aaf220 715
828c7908
BW		 716 i915_check_and_clear_faults(dev);
717
bee4a186 718 /* First fill our portion of the GTT with scratch pages */
853ba5d2
BW		 719 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
720 dev_priv->gtt.base.start / PAGE_SIZE,
828c7908
BW		 721 dev_priv->gtt.base.total / PAGE_SIZE,
722 true);
bee4a186 723
35c20a60 724 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2c22569b 725 i915_gem_clflush_object(obj, obj->pin_display);
74163907 726 i915_gem_gtt_bind_object(obj, obj->cache_level);
76aaf220
DV		 727 }
728
e76e9aeb 729 i915_gem_chipset_flush(dev);
76aaf220 730}
7c2e6fdf 731
74163907 732int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
7c2e6fdf 733{
9da3da66 734 if (obj->has_dma_mapping)
74163907 735 return 0;
9da3da66
CW		 736
737 if (!dma_map_sg(&obj->base.dev->pdev->dev,
738 obj->pages->sgl, obj->pages->nents,
739 PCI_DMA_BIDIRECTIONAL))
740 return -ENOSPC;
741
742 return 0;
7c2e6fdf
DV		 743}
744
94ec8f61
BW		 745static inline void gen8_set_pte(void __iomem *addr, gen8_gtt_pte_t pte)
746{
747#ifdef writeq
748 writeq(pte, addr);
749#else
750 iowrite32((u32)pte, addr);
751 iowrite32(pte >> 32, addr + 4);
752#endif
753}
754
755static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
756 struct sg_table *st,
757 unsigned int first_entry,
758 enum i915_cache_level level)
759{
760 struct drm_i915_private *dev_priv = vm->dev->dev_private;
761 gen8_gtt_pte_t __iomem *gtt_entries =
762 (gen8_gtt_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
763 int i = 0;
764 struct sg_page_iter sg_iter;
765 dma_addr_t addr;
766
767 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
768 addr = sg_dma_address(sg_iter.sg) +
769 (sg_iter.sg_pgoffset << PAGE_SHIFT);
770 gen8_set_pte(&gtt_entries[i],
771 gen8_pte_encode(addr, level, true));
772 i++;
773 }
774
775 /*
776 * XXX: This serves as a posting read to make sure that the PTE has
777 * actually been updated. There is some concern that even though
778 * registers and PTEs are within the same BAR that they are potentially
779 * of NUMA access patterns. Therefore, even with the way we assume
780 * hardware should work, we must keep this posting read for paranoia.
781 */
782 if (i != 0)
783 WARN_ON(readq(&gtt_entries[i-1])
784 != gen8_pte_encode(addr, level, true));
785
786#if 0 /* TODO: Still needed on GEN8? */
787 /* This next bit makes the above posting read even more important. We
788 * want to flush the TLBs only after we're certain all the PTE updates
789 * have finished.
790 */
791 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
792 POSTING_READ(GFX_FLSH_CNTL_GEN6);
793#endif
794}
795
e76e9aeb
BW		 796/*
797 * Binds an object into the global gtt with the specified cache level. The object
798 * will be accessible to the GPU via commands whose operands reference offsets
799 * within the global GTT as well as accessible by the GPU through the GMADR
800 * mapped BAR (dev_priv->mm.gtt->gtt).
801 */
853ba5d2 802static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
7faf1ab2
DV		 803 struct sg_table *st,
804 unsigned int first_entry,
805 enum i915_cache_level level)
e76e9aeb 806{
853ba5d2 807 struct drm_i915_private *dev_priv = vm->dev->dev_private;
e7c2b58b
BW		 808 gen6_gtt_pte_t __iomem *gtt_entries =
809 (gen6_gtt_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
6e995e23
ID		 810 int i = 0;
811 struct sg_page_iter sg_iter;
e76e9aeb
BW		 812 dma_addr_t addr;
813
6e995e23 814 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
2db76d7c 815 addr = sg_page_iter_dma_address(&sg_iter);
b35b380e 816 iowrite32(vm->pte_encode(addr, level, true), &gtt_entries[i]);
6e995e23 817 i++;
e76e9aeb
BW		 818 }
819
e76e9aeb
BW		 820 /* XXX: This serves as a posting read to make sure that the PTE has
821 * actually been updated. There is some concern that even though
822 * registers and PTEs are within the same BAR that they are potentially
823 * of NUMA access patterns. Therefore, even with the way we assume
824 * hardware should work, we must keep this posting read for paranoia.
825 */
826 if (i != 0)
853ba5d2 827 WARN_ON(readl(&gtt_entries[i-1]) !=
b35b380e 828 vm->pte_encode(addr, level, true));
0f9b91c7
BW		 829
830 /* This next bit makes the above posting read even more important. We
831 * want to flush the TLBs only after we're certain all the PTE updates
832 * have finished.
833 */
834 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
835 POSTING_READ(GFX_FLSH_CNTL_GEN6);
e76e9aeb
BW		 836}
837
94ec8f61
BW		 838static void gen8_ggtt_clear_range(struct i915_address_space *vm,
839 unsigned int first_entry,
840 unsigned int num_entries,
841 bool use_scratch)
842{
843 struct drm_i915_private *dev_priv = vm->dev->dev_private;
844 gen8_gtt_pte_t scratch_pte, __iomem *gtt_base =
845 (gen8_gtt_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
846 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
847 int i;
848
849 if (WARN(num_entries > max_entries,
850 "First entry = %d; Num entries = %d (max=%d)\n",
851 first_entry, num_entries, max_entries))
852 num_entries = max_entries;
853
854 scratch_pte = gen8_pte_encode(vm->scratch.addr,
855 I915_CACHE_LLC,
856 use_scratch);
857 for (i = 0; i < num_entries; i++)
858 gen8_set_pte(&gtt_base[i], scratch_pte);
859 readl(gtt_base);
860}
861
853ba5d2 862static void gen6_ggtt_clear_range(struct i915_address_space *vm,
7faf1ab2 863 unsigned int first_entry,
828c7908
BW		 864 unsigned int num_entries,
865 bool use_scratch)
7faf1ab2 866{
853ba5d2 867 struct drm_i915_private *dev_priv = vm->dev->dev_private;
e7c2b58b
BW		 868 gen6_gtt_pte_t scratch_pte, __iomem *gtt_base =
869 (gen6_gtt_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
a54c0c27 870 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
7faf1ab2
DV		 871 int i;
872
873 if (WARN(num_entries > max_entries,
874 "First entry = %d; Num entries = %d (max=%d)\n",
875 first_entry, num_entries, max_entries))
876 num_entries = max_entries;
877
828c7908
BW		 878 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, use_scratch);
879
7faf1ab2
DV		 880 for (i = 0; i < num_entries; i++)
881 iowrite32(scratch_pte, &gtt_base[i]);
882 readl(gtt_base);
883}
884
853ba5d2 885static void i915_ggtt_insert_entries(struct i915_address_space *vm,
7faf1ab2
DV		 886 struct sg_table *st,
887 unsigned int pg_start,
888 enum i915_cache_level cache_level)
889{
890 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
891 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
892
893 intel_gtt_insert_sg_entries(st, pg_start, flags);
894
895}
896
853ba5d2 897static void i915_ggtt_clear_range(struct i915_address_space *vm,
7faf1ab2 898 unsigned int first_entry,
828c7908
BW		 899 unsigned int num_entries,
900 bool unused)
7faf1ab2
DV		 901{
902 intel_gtt_clear_range(first_entry, num_entries);
903}
904
905
74163907
DV		 906void i915_gem_gtt_bind_object(struct drm_i915_gem_object *obj,
907 enum i915_cache_level cache_level)
d5bd1449
CW		 908{
909 struct drm_device *dev = obj->base.dev;
7faf1ab2 910 struct drm_i915_private *dev_priv = dev->dev_private;
853ba5d2 911 const unsigned long entry = i915_gem_obj_ggtt_offset(obj) >> PAGE_SHIFT;
7faf1ab2 912
853ba5d2
BW		 913 dev_priv->gtt.base.insert_entries(&dev_priv->gtt.base, obj->pages,
914 entry,
915 cache_level);
d5bd1449 916
74898d7e 917 obj->has_global_gtt_mapping = 1;
d5bd1449
CW		 918}
919
05394f39 920void i915_gem_gtt_unbind_object(struct drm_i915_gem_object *obj)
74163907 921{
7faf1ab2
DV		 922 struct drm_device *dev = obj->base.dev;
923 struct drm_i915_private *dev_priv = dev->dev_private;
853ba5d2 924 const unsigned long entry = i915_gem_obj_ggtt_offset(obj) >> PAGE_SHIFT;
7faf1ab2 925
853ba5d2
BW		 926 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
927 entry,
828c7908
BW		 928 obj->base.size >> PAGE_SHIFT,
929 true);
74898d7e
DV		 930
931 obj->has_global_gtt_mapping = 0;
74163907
DV		 932}
933
934void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
7c2e6fdf 935{
5c042287
BW		 936 struct drm_device *dev = obj->base.dev;
937 struct drm_i915_private *dev_priv = dev->dev_private;
938 bool interruptible;
939
940 interruptible = do_idling(dev_priv);
941
9da3da66
CW		 942 if (!obj->has_dma_mapping)
943 dma_unmap_sg(&dev->pdev->dev,
944 obj->pages->sgl, obj->pages->nents,
945 PCI_DMA_BIDIRECTIONAL);
5c042287
BW		 946
947 undo_idling(dev_priv, interruptible);
7c2e6fdf 948}
644ec02b 949
42d6ab48
CW		 950static void i915_gtt_color_adjust(struct drm_mm_node *node,
951 unsigned long color,
952 unsigned long *start,
953 unsigned long *end)
954{
955 if (node->color != color)
956 *start += 4096;
957
958 if (!list_empty(&node->node_list)) {
959 node = list_entry(node->node_list.next,
960 struct drm_mm_node,
961 node_list);
962 if (node->allocated && node->color != color)
963 *end -= 4096;
964 }
965}
fbe5d36e 966
d7e5008f
BW		 967void i915_gem_setup_global_gtt(struct drm_device *dev,
968 unsigned long start,
969 unsigned long mappable_end,
970 unsigned long end)
644ec02b 971{
e78891ca
BW		 972 /* Let GEM Manage all of the aperture.
973 *
974 * However, leave one page at the end still bound to the scratch page.
975 * There are a number of places where the hardware apparently prefetches
976 * past the end of the object, and we've seen multiple hangs with the
977 * GPU head pointer stuck in a batchbuffer bound at the last page of the
978 * aperture. One page should be enough to keep any prefetching inside
979 * of the aperture.
980 */
40d74980
BW		 981 struct drm_i915_private *dev_priv = dev->dev_private;
982 struct i915_address_space *ggtt_vm = &dev_priv->gtt.base;
ed2f3452
CW		 983 struct drm_mm_node *entry;
984 struct drm_i915_gem_object *obj;
985 unsigned long hole_start, hole_end;
644ec02b 986
35451cb6
BW		 987 BUG_ON(mappable_end > end);
988
ed2f3452 989 /* Subtract the guard page ... */
40d74980 990 drm_mm_init(&ggtt_vm->mm, start, end - start - PAGE_SIZE);
42d6ab48 991 if (!HAS_LLC(dev))
93bd8649 992 dev_priv->gtt.base.mm.color_adjust = i915_gtt_color_adjust;
644ec02b 993
ed2f3452 994 /* Mark any preallocated objects as occupied */
35c20a60 995 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
40d74980 996 struct i915_vma *vma = i915_gem_obj_to_vma(obj, ggtt_vm);
b3a070cc 997 int ret;
edd41a87 998 DRM_DEBUG_KMS("reserving preallocated space: %lx + %zx\n",
c6cfb325
BW		 999 i915_gem_obj_ggtt_offset(obj), obj->base.size);
1000
1001 WARN_ON(i915_gem_obj_ggtt_bound(obj));
40d74980 1002 ret = drm_mm_reserve_node(&ggtt_vm->mm, &vma->node);
c6cfb325 1003 if (ret)
b3a070cc 1004 DRM_DEBUG_KMS("Reservation failed\n");
ed2f3452 1005 obj->has_global_gtt_mapping = 1;
2f633156 1006 list_add(&vma->vma_link, &obj->vma_list);
ed2f3452
CW		 1007 }
1008
853ba5d2
BW		 1009 dev_priv->gtt.base.start = start;
1010 dev_priv->gtt.base.total = end - start;
644ec02b 1011
ed2f3452 1012 /* Clear any non-preallocated blocks */
40d74980 1013 drm_mm_for_each_hole(entry, &ggtt_vm->mm, hole_start, hole_end) {
853ba5d2 1014 const unsigned long count = (hole_end - hole_start) / PAGE_SIZE;
ed2f3452
CW		 1015 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
1016 hole_start, hole_end);
828c7908 1017 ggtt_vm->clear_range(ggtt_vm, hole_start / PAGE_SIZE, count, true);
ed2f3452
CW		 1018 }
1019
1020 /* And finally clear the reserved guard page */
828c7908 1021 ggtt_vm->clear_range(ggtt_vm, end / PAGE_SIZE - 1, 1, true);
e76e9aeb
BW		 1022}
1023
d7e5008f
BW		 1024static bool
1025intel_enable_ppgtt(struct drm_device *dev)
1026{
1027 if (i915_enable_ppgtt >= 0)
1028 return i915_enable_ppgtt;
1029
1030#ifdef CONFIG_INTEL_IOMMU
1031 /* Disable ppgtt on SNB if VT-d is on. */
1032 if (INTEL_INFO(dev)->gen == 6 && intel_iommu_gfx_mapped)
1033 return false;
1034#endif
1035
1036 return true;
1037}
1038
1039void i915_gem_init_global_gtt(struct drm_device *dev)
1040{
1041 struct drm_i915_private *dev_priv = dev->dev_private;
1042 unsigned long gtt_size, mappable_size;
d7e5008f 1043
853ba5d2 1044 gtt_size = dev_priv->gtt.base.total;
93d18799 1045 mappable_size = dev_priv->gtt.mappable_end;
d7e5008f
BW		 1046
1047 if (intel_enable_ppgtt(dev) && HAS_ALIASING_PPGTT(dev)) {
e78891ca 1048 int ret;
3eb1c005
BW		 1049
1050 if (INTEL_INFO(dev)->gen <= 7) {
1051 /* PPGTT pdes are stolen from global gtt ptes, so shrink the
1052 * aperture accordingly when using aliasing ppgtt. */
6670a5a5 1053 gtt_size -= GEN6_PPGTT_PD_ENTRIES * PAGE_SIZE;
3eb1c005 1054 }
d7e5008f
BW		 1055
1056 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
1057
1058 ret = i915_gem_init_aliasing_ppgtt(dev);
e78891ca 1059 if (!ret)
d7e5008f 1060 return;
e78891ca
BW		 1061
1062 DRM_ERROR("Aliased PPGTT setup failed %d\n", ret);
93bd8649 1063 drm_mm_takedown(&dev_priv->gtt.base.mm);
6670a5a5 1064 gtt_size += GEN6_PPGTT_PD_ENTRIES * PAGE_SIZE;
d7e5008f 1065 }
e78891ca 1066 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
e76e9aeb
BW		 1067}
1068
1069static int setup_scratch_page(struct drm_device *dev)
1070{
1071 struct drm_i915_private *dev_priv = dev->dev_private;
1072 struct page *page;
1073 dma_addr_t dma_addr;
1074
1075 page = alloc_page(GFP_KERNEL | GFP_DMA32 | __GFP_ZERO);
1076 if (page == NULL)
1077 return -ENOMEM;
1078 get_page(page);
1079 set_pages_uc(page, 1);
1080
1081#ifdef CONFIG_INTEL_IOMMU
1082 dma_addr = pci_map_page(dev->pdev, page, 0, PAGE_SIZE,
1083 PCI_DMA_BIDIRECTIONAL);
1084 if (pci_dma_mapping_error(dev->pdev, dma_addr))
1085 return -EINVAL;
1086#else
1087 dma_addr = page_to_phys(page);
1088#endif
853ba5d2
BW		 1089 dev_priv->gtt.base.scratch.page = page;
1090 dev_priv->gtt.base.scratch.addr = dma_addr;
e76e9aeb
BW		 1091
1092 return 0;
1093}
1094
1095static void teardown_scratch_page(struct drm_device *dev)
1096{
1097 struct drm_i915_private *dev_priv = dev->dev_private;
853ba5d2
BW		 1098 struct page *page = dev_priv->gtt.base.scratch.page;
1099
1100 set_pages_wb(page, 1);
1101 pci_unmap_page(dev->pdev, dev_priv->gtt.base.scratch.addr,
e76e9aeb 1102 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
853ba5d2
BW		 1103 put_page(page);
1104 __free_page(page);
e76e9aeb
BW		 1105}
1106
1107static inline unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
1108{
1109 snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
1110 snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
1111 return snb_gmch_ctl << 20;
1112}
1113
9459d252
BW		 1114static inline unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
1115{
1116 bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
1117 bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
1118 if (bdw_gmch_ctl)
1119 bdw_gmch_ctl = 1 << bdw_gmch_ctl;
1120 return bdw_gmch_ctl << 20;
1121}
1122
baa09f5f 1123static inline size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
e76e9aeb
BW		 1124{
1125 snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
1126 snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
1127 return snb_gmch_ctl << 25; /* 32 MB units */
1128}
1129
9459d252
BW		 1130static inline size_t gen8_get_stolen_size(u16 bdw_gmch_ctl)
1131{
1132 bdw_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
1133 bdw_gmch_ctl &= BDW_GMCH_GMS_MASK;
1134 return bdw_gmch_ctl << 25; /* 32 MB units */
1135}
1136
63340133
BW		 1137static int ggtt_probe_common(struct drm_device *dev,
1138 size_t gtt_size)
1139{
1140 struct drm_i915_private *dev_priv = dev->dev_private;
1141 phys_addr_t gtt_bus_addr;
1142 int ret;
1143
1144 /* For Modern GENs the PTEs and register space are split in the BAR */
1145 gtt_bus_addr = pci_resource_start(dev->pdev, 0) +
1146 (pci_resource_len(dev->pdev, 0) / 2);
1147
1148 dev_priv->gtt.gsm = ioremap_wc(gtt_bus_addr, gtt_size);
1149 if (!dev_priv->gtt.gsm) {
1150 DRM_ERROR("Failed to map the gtt page table\n");
1151 return -ENOMEM;
1152 }
1153
1154 ret = setup_scratch_page(dev);
1155 if (ret) {
1156 DRM_ERROR("Scratch setup failed\n");
1157 /* iounmap will also get called at remove, but meh */
1158 iounmap(dev_priv->gtt.gsm);
1159 }
1160
1161 return ret;
1162}
1163
fbe5d36e
BW		 1164/* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
1165 * bits. When using advanced contexts each context stores its own PAT, but
1166 * writing this data shouldn't be harmful even in those cases. */
1167static void gen8_setup_private_ppat(struct drm_i915_private *dev_priv)
1168{
1169#define GEN8_PPAT_UC (0<<0)
1170#define GEN8_PPAT_WC (1<<0)
1171#define GEN8_PPAT_WT (2<<0)
1172#define GEN8_PPAT_WB (3<<0)
1173#define GEN8_PPAT_ELLC_OVERRIDE (0<<2)
1174/* FIXME(BDW): Bspec is completely confused about cache control bits. */
1175#define GEN8_PPAT_LLC (1<<2)
1176#define GEN8_PPAT_LLCELLC (2<<2)
1177#define GEN8_PPAT_LLCeLLC (3<<2)
1178#define GEN8_PPAT_AGE(x) (x<<4)
1179#define GEN8_PPAT(i, x) ((uint64_t) (x) << ((i) * 8))
1180 uint64_t pat;
1181
1182 pat = GEN8_PPAT(0, GEN8_PPAT_WB | GEN8_PPAT_LLC) | /* for normal objects, no eLLC */
1183 GEN8_PPAT(1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC) | /* for something pointing to ptes? */
1184 GEN8_PPAT(2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC) | /* for scanout with eLLC */
1185 GEN8_PPAT(3, GEN8_PPAT_UC) | /* Uncached objects, mostly for scanout */
1186 GEN8_PPAT(4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0)) |
1187 GEN8_PPAT(5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1)) |
1188 GEN8_PPAT(6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2)) |
1189 GEN8_PPAT(7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
1190
1191 /* XXX: spec defines this as 2 distinct registers. It's unclear if a 64b
1192 * write would work. */
1193 I915_WRITE(GEN8_PRIVATE_PAT, pat);
1194 I915_WRITE(GEN8_PRIVATE_PAT + 4, pat >> 32);
1195}
1196
63340133
BW		 1197static int gen8_gmch_probe(struct drm_device *dev,
1198 size_t *gtt_total,
1199 size_t *stolen,
1200 phys_addr_t *mappable_base,
1201 unsigned long *mappable_end)
1202{
1203 struct drm_i915_private *dev_priv = dev->dev_private;
1204 unsigned int gtt_size;
1205 u16 snb_gmch_ctl;
1206 int ret;
1207
1208 /* TODO: We're not aware of mappable constraints on gen8 yet */
1209 *mappable_base = pci_resource_start(dev->pdev, 2);
1210 *mappable_end = pci_resource_len(dev->pdev, 2);
1211
1212 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(39)))
1213 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(39));
1214
1215 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
1216
1217 *stolen = gen8_get_stolen_size(snb_gmch_ctl);
1218
1219 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
d31eb10e 1220 *gtt_total = (gtt_size / sizeof(gen8_gtt_pte_t)) << PAGE_SHIFT;
63340133 1221
fbe5d36e
BW		 1222 gen8_setup_private_ppat(dev_priv);
1223
63340133
BW		 1224 ret = ggtt_probe_common(dev, gtt_size);
1225
94ec8f61
BW		 1226 dev_priv->gtt.base.clear_range = gen8_ggtt_clear_range;
1227 dev_priv->gtt.base.insert_entries = gen8_ggtt_insert_entries;
63340133
BW		 1228
1229 return ret;
1230}
1231
baa09f5f
BW		 1232static int gen6_gmch_probe(struct drm_device *dev,
1233 size_t *gtt_total,
41907ddc
BW		 1234 size_t *stolen,
1235 phys_addr_t *mappable_base,
1236 unsigned long *mappable_end)
e76e9aeb
BW		 1237{
1238 struct drm_i915_private *dev_priv = dev->dev_private;
baa09f5f 1239 unsigned int gtt_size;
e76e9aeb 1240 u16 snb_gmch_ctl;
e76e9aeb
BW		 1241 int ret;
1242
41907ddc
BW		 1243 *mappable_base = pci_resource_start(dev->pdev, 2);
1244 *mappable_end = pci_resource_len(dev->pdev, 2);
1245
baa09f5f
BW		 1246 /* 64/512MB is the current min/max we actually know of, but this is just
1247 * a coarse sanity check.
e76e9aeb 1248 */
41907ddc 1249 if ((*mappable_end < (64<<20) || (*mappable_end > (512<<20)))) {
baa09f5f
BW		 1250 DRM_ERROR("Unknown GMADR size (%lx)\n",
1251 dev_priv->gtt.mappable_end);
1252 return -ENXIO;
e76e9aeb
BW		 1253 }
1254
e76e9aeb
BW		 1255 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
1256 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
e76e9aeb 1257 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
e76e9aeb 1258
63340133 1259 *stolen = gen6_get_stolen_size(snb_gmch_ctl);
e76e9aeb 1260
63340133
BW		 1261 gtt_size = gen6_get_total_gtt_size(snb_gmch_ctl);
1262 *gtt_total = (gtt_size / sizeof(gen6_gtt_pte_t)) << PAGE_SHIFT;
a93e4161 1263
63340133 1264 ret = ggtt_probe_common(dev, gtt_size);
e76e9aeb 1265
853ba5d2
BW		 1266 dev_priv->gtt.base.clear_range = gen6_ggtt_clear_range;
1267 dev_priv->gtt.base.insert_entries = gen6_ggtt_insert_entries;
7faf1ab2 1268
e76e9aeb
BW		 1269 return ret;
1270}
1271
853ba5d2 1272static void gen6_gmch_remove(struct i915_address_space *vm)
e76e9aeb 1273{
853ba5d2
BW		 1274
1275 struct i915_gtt *gtt = container_of(vm, struct i915_gtt, base);
1276 iounmap(gtt->gsm);
1277 teardown_scratch_page(vm->dev);
644ec02b 1278}
baa09f5f
BW		 1279
1280static int i915_gmch_probe(struct drm_device *dev,
1281 size_t *gtt_total,
41907ddc
BW		 1282 size_t *stolen,
1283 phys_addr_t *mappable_base,
1284 unsigned long *mappable_end)
baa09f5f
BW		 1285{
1286 struct drm_i915_private *dev_priv = dev->dev_private;
1287 int ret;
1288
baa09f5f
BW		 1289 ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL);
1290 if (!ret) {
1291 DRM_ERROR("failed to set up gmch\n");
1292 return -EIO;
1293 }
1294
41907ddc 1295 intel_gtt_get(gtt_total, stolen, mappable_base, mappable_end);
baa09f5f
BW		 1296
1297 dev_priv->gtt.do_idle_maps = needs_idle_maps(dev_priv->dev);
853ba5d2
BW		 1298 dev_priv->gtt.base.clear_range = i915_ggtt_clear_range;
1299 dev_priv->gtt.base.insert_entries = i915_ggtt_insert_entries;
baa09f5f
BW		 1300
1301 return 0;
1302}
1303
853ba5d2 1304static void i915_gmch_remove(struct i915_address_space *vm)
baa09f5f
BW		 1305{
1306 intel_gmch_remove();
1307}
1308
1309int i915_gem_gtt_init(struct drm_device *dev)
1310{
1311 struct drm_i915_private *dev_priv = dev->dev_private;
1312 struct i915_gtt *gtt = &dev_priv->gtt;
baa09f5f
BW		 1313 int ret;
1314
baa09f5f 1315 if (INTEL_INFO(dev)->gen <= 5) {
b2f21b4d 1316 gtt->gtt_probe = i915_gmch_probe;
853ba5d2 1317 gtt->base.cleanup = i915_gmch_remove;
63340133 1318 } else if (INTEL_INFO(dev)->gen < 8) {
b2f21b4d 1319 gtt->gtt_probe = gen6_gmch_probe;
853ba5d2 1320 gtt->base.cleanup = gen6_gmch_remove;
4d15c145 1321 if (IS_HASWELL(dev) && dev_priv->ellc_size)
853ba5d2 1322 gtt->base.pte_encode = iris_pte_encode;
4d15c145 1323 else if (IS_HASWELL(dev))
853ba5d2 1324 gtt->base.pte_encode = hsw_pte_encode;
b2f21b4d 1325 else if (IS_VALLEYVIEW(dev))
853ba5d2 1326 gtt->base.pte_encode = byt_pte_encode;
350ec881
CW		 1327 else if (INTEL_INFO(dev)->gen >= 7)
1328 gtt->base.pte_encode = ivb_pte_encode;
b2f21b4d 1329 else
350ec881 1330 gtt->base.pte_encode = snb_pte_encode;
63340133
BW		 1331 } else {
1332 dev_priv->gtt.gtt_probe = gen8_gmch_probe;
1333 dev_priv->gtt.base.cleanup = gen6_gmch_remove;
baa09f5f
BW		 1334 }
1335
853ba5d2 1336 ret = gtt->gtt_probe(dev, &gtt->base.total, &gtt->stolen_size,
b2f21b4d 1337 &gtt->mappable_base, &gtt->mappable_end);
a54c0c27 1338 if (ret)
baa09f5f 1339 return ret;
baa09f5f 1340
853ba5d2
BW		 1341 gtt->base.dev = dev;
1342
baa09f5f 1343 /* GMADR is the PCI mmio aperture into the global GTT. */
853ba5d2
BW		 1344 DRM_INFO("Memory usable by graphics device = %zdM\n",
1345 gtt->base.total >> 20);
b2f21b4d
BW		 1346 DRM_DEBUG_DRIVER("GMADR size = %ldM\n", gtt->mappable_end >> 20);
1347 DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", gtt->stolen_size >> 20);
baa09f5f
BW		 1348
1349 return 0;
1350}
Linux 6.x block layer and io_uring tree(s)