?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge branch 'akpm' (patches from Andrew)
[linux/fpc-iii.git] / drivers / gpu / drm / i915 / gvt / gtt.c
blob34cb404ba4b789ca97c86794d90c24edf3cf95ef
1 /*
2 * GTT virtualization
3 *
4 * Copyright(c) 2011-2016 Intel Corporation. All rights reserved.
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 FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 * SOFTWARE.
24 *
25 * Authors:
26 * Zhi Wang <zhi.a.wang@intel.com>
27 * Zhenyu Wang <zhenyuw@linux.intel.com>
28 * Xiao Zheng <xiao.zheng@intel.com>
29 *
30 * Contributors:
31 * Min He <min.he@intel.com>
32 * Bing Niu <bing.niu@intel.com>
33 *
34 */
35
36 #include "i915_drv.h"
37 #include "gvt.h"
38 #include "i915_pvinfo.h"
39 #include "trace.h"
40
41 #if defined(VERBOSE_DEBUG)
42 #define gvt_vdbg_mm(fmt, args...) gvt_dbg_mm(fmt, ##args)
43 #else
44 #define gvt_vdbg_mm(fmt, args...)
45 #endif
46
47 static bool enable_out_of_sync = false;
48 static int preallocated_oos_pages = 8192;
49
50 /*
51 * validate a gm address and related range size,
52 * translate it to host gm address
53 */
54 bool intel_gvt_ggtt_validate_range(struct intel_vgpu *vgpu, u64 addr, u32 size)
55 {
56 if (size == 0)
57 return vgpu_gmadr_is_valid(vgpu, addr);
58
59 if (vgpu_gmadr_is_aperture(vgpu, addr) &&
60 vgpu_gmadr_is_aperture(vgpu, addr + size - 1))
61 return true;
62 else if (vgpu_gmadr_is_hidden(vgpu, addr) &&
63 vgpu_gmadr_is_hidden(vgpu, addr + size - 1))
64 return true;
65
66 gvt_dbg_mm("Invalid ggtt range at 0x%llx, size: 0x%x\n",
67 addr, size);
68 return false;
69 }
70
71 /* translate a guest gmadr to host gmadr */
72 int intel_gvt_ggtt_gmadr_g2h(struct intel_vgpu *vgpu, u64 g_addr, u64 *h_addr)
73 {
74 if (WARN(!vgpu_gmadr_is_valid(vgpu, g_addr),
75 "invalid guest gmadr %llx\n", g_addr))
76 return -EACCES;
77
78 if (vgpu_gmadr_is_aperture(vgpu, g_addr))
79 *h_addr = vgpu_aperture_gmadr_base(vgpu)
80 + (g_addr - vgpu_aperture_offset(vgpu));
81 else
82 *h_addr = vgpu_hidden_gmadr_base(vgpu)
83 + (g_addr - vgpu_hidden_offset(vgpu));
84 return 0;
85 }
86
87 /* translate a host gmadr to guest gmadr */
88 int intel_gvt_ggtt_gmadr_h2g(struct intel_vgpu *vgpu, u64 h_addr, u64 *g_addr)
89 {
90 if (WARN(!gvt_gmadr_is_valid(vgpu->gvt, h_addr),
91 "invalid host gmadr %llx\n", h_addr))
92 return -EACCES;
93
94 if (gvt_gmadr_is_aperture(vgpu->gvt, h_addr))
95 *g_addr = vgpu_aperture_gmadr_base(vgpu)
96 + (h_addr - gvt_aperture_gmadr_base(vgpu->gvt));
97 else
98 *g_addr = vgpu_hidden_gmadr_base(vgpu)
99 + (h_addr - gvt_hidden_gmadr_base(vgpu->gvt));
100 return 0;
101 }
102
103 int intel_gvt_ggtt_index_g2h(struct intel_vgpu *vgpu, unsigned long g_index,
104 unsigned long *h_index)
105 {
106 u64 h_addr;
107 int ret;
108
109 ret = intel_gvt_ggtt_gmadr_g2h(vgpu, g_index << I915_GTT_PAGE_SHIFT,
110 &h_addr);
111 if (ret)
112 return ret;
113
114 *h_index = h_addr >> I915_GTT_PAGE_SHIFT;
115 return 0;
116 }
117
118 int intel_gvt_ggtt_h2g_index(struct intel_vgpu *vgpu, unsigned long h_index,
119 unsigned long *g_index)
120 {
121 u64 g_addr;
122 int ret;
123
124 ret = intel_gvt_ggtt_gmadr_h2g(vgpu, h_index << I915_GTT_PAGE_SHIFT,
125 &g_addr);
126 if (ret)
127 return ret;
128
129 *g_index = g_addr >> I915_GTT_PAGE_SHIFT;
130 return 0;
131 }
132
133 #define gtt_type_is_entry(type) \
134 (type > GTT_TYPE_INVALID && type < GTT_TYPE_PPGTT_ENTRY \
135 && type != GTT_TYPE_PPGTT_PTE_ENTRY \
136 && type != GTT_TYPE_PPGTT_ROOT_ENTRY)
137
138 #define gtt_type_is_pt(type) \
139 (type >= GTT_TYPE_PPGTT_PTE_PT && type < GTT_TYPE_MAX)
140
141 #define gtt_type_is_pte_pt(type) \
142 (type == GTT_TYPE_PPGTT_PTE_PT)
143
144 #define gtt_type_is_root_pointer(type) \
145 (gtt_type_is_entry(type) && type > GTT_TYPE_PPGTT_ROOT_ENTRY)
146
147 #define gtt_init_entry(e, t, p, v) do { \
148 (e)->type = t; \
149 (e)->pdev = p; \
150 memcpy(&(e)->val64, &v, sizeof(v)); \
151 } while (0)
152
153 /*
154 * Mappings between GTT_TYPE* enumerations.
155 * Following information can be found according to the given type:
156 * - type of next level page table
157 * - type of entry inside this level page table
158 * - type of entry with PSE set
159 *
160 * If the given type doesn't have such a kind of information,
161 * e.g. give a l4 root entry type, then request to get its PSE type,
162 * give a PTE page table type, then request to get its next level page
163 * table type, as we know l4 root entry doesn't have a PSE bit,
164 * and a PTE page table doesn't have a next level page table type,
165 * GTT_TYPE_INVALID will be returned. This is useful when traversing a
166 * page table.
167 */
168
169 struct gtt_type_table_entry {
170 int entry_type;
171 int pt_type;
172 int next_pt_type;
173 int pse_entry_type;
174 };
175
176 #define GTT_TYPE_TABLE_ENTRY(type, e_type, cpt_type, npt_type, pse_type) \
177 [type] = { \
178 .entry_type = e_type, \
179 .pt_type = cpt_type, \
180 .next_pt_type = npt_type, \
181 .pse_entry_type = pse_type, \
182 }
183
184 static struct gtt_type_table_entry gtt_type_table[] = {
185 GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_ROOT_L4_ENTRY,
186 GTT_TYPE_PPGTT_ROOT_L4_ENTRY,
187 GTT_TYPE_INVALID,
188 GTT_TYPE_PPGTT_PML4_PT,
189 GTT_TYPE_INVALID),
190 GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PML4_PT,
191 GTT_TYPE_PPGTT_PML4_ENTRY,
192 GTT_TYPE_PPGTT_PML4_PT,
193 GTT_TYPE_PPGTT_PDP_PT,
194 GTT_TYPE_INVALID),
195 GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PML4_ENTRY,
196 GTT_TYPE_PPGTT_PML4_ENTRY,
197 GTT_TYPE_PPGTT_PML4_PT,
198 GTT_TYPE_PPGTT_PDP_PT,
199 GTT_TYPE_INVALID),
200 GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PDP_PT,
201 GTT_TYPE_PPGTT_PDP_ENTRY,
202 GTT_TYPE_PPGTT_PDP_PT,
203 GTT_TYPE_PPGTT_PDE_PT,
204 GTT_TYPE_PPGTT_PTE_1G_ENTRY),
205 GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_ROOT_L3_ENTRY,
206 GTT_TYPE_PPGTT_ROOT_L3_ENTRY,
207 GTT_TYPE_INVALID,
208 GTT_TYPE_PPGTT_PDE_PT,
209 GTT_TYPE_PPGTT_PTE_1G_ENTRY),
210 GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PDP_ENTRY,
211 GTT_TYPE_PPGTT_PDP_ENTRY,
212 GTT_TYPE_PPGTT_PDP_PT,
213 GTT_TYPE_PPGTT_PDE_PT,
214 GTT_TYPE_PPGTT_PTE_1G_ENTRY),
215 GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PDE_PT,
216 GTT_TYPE_PPGTT_PDE_ENTRY,
217 GTT_TYPE_PPGTT_PDE_PT,
218 GTT_TYPE_PPGTT_PTE_PT,
219 GTT_TYPE_PPGTT_PTE_2M_ENTRY),
220 GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PDE_ENTRY,
221 GTT_TYPE_PPGTT_PDE_ENTRY,
222 GTT_TYPE_PPGTT_PDE_PT,
223 GTT_TYPE_PPGTT_PTE_PT,
224 GTT_TYPE_PPGTT_PTE_2M_ENTRY),
225 /* We take IPS bit as 'PSE' for PTE level. */
226 GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PTE_PT,
227 GTT_TYPE_PPGTT_PTE_4K_ENTRY,
228 GTT_TYPE_PPGTT_PTE_PT,
229 GTT_TYPE_INVALID,
230 GTT_TYPE_PPGTT_PTE_64K_ENTRY),
231 GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PTE_4K_ENTRY,
232 GTT_TYPE_PPGTT_PTE_4K_ENTRY,
233 GTT_TYPE_PPGTT_PTE_PT,
234 GTT_TYPE_INVALID,
235 GTT_TYPE_PPGTT_PTE_64K_ENTRY),
236 GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PTE_64K_ENTRY,
237 GTT_TYPE_PPGTT_PTE_4K_ENTRY,
238 GTT_TYPE_PPGTT_PTE_PT,
239 GTT_TYPE_INVALID,
240 GTT_TYPE_PPGTT_PTE_64K_ENTRY),
241 GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PTE_2M_ENTRY,
242 GTT_TYPE_PPGTT_PDE_ENTRY,
243 GTT_TYPE_PPGTT_PDE_PT,
244 GTT_TYPE_INVALID,
245 GTT_TYPE_PPGTT_PTE_2M_ENTRY),
246 GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PTE_1G_ENTRY,
247 GTT_TYPE_PPGTT_PDP_ENTRY,
248 GTT_TYPE_PPGTT_PDP_PT,
249 GTT_TYPE_INVALID,
250 GTT_TYPE_PPGTT_PTE_1G_ENTRY),
251 GTT_TYPE_TABLE_ENTRY(GTT_TYPE_GGTT_PTE,
252 GTT_TYPE_GGTT_PTE,
253 GTT_TYPE_INVALID,
254 GTT_TYPE_INVALID,
255 GTT_TYPE_INVALID),
256 };
257
258 static inline int get_next_pt_type(int type)
259 {
260 return gtt_type_table[type].next_pt_type;
261 }
262
263 static inline int get_pt_type(int type)
264 {
265 return gtt_type_table[type].pt_type;
266 }
267
268 static inline int get_entry_type(int type)
269 {
270 return gtt_type_table[type].entry_type;
271 }
272
273 static inline int get_pse_type(int type)
274 {
275 return gtt_type_table[type].pse_entry_type;
276 }
277
278 static u64 read_pte64(struct drm_i915_private *dev_priv, unsigned long index)
279 {
280 void __iomem *addr = (gen8_pte_t __iomem *)dev_priv->ggtt.gsm + index;
281
282 return readq(addr);
283 }
284
285 static void ggtt_invalidate(struct drm_i915_private *dev_priv)
286 {
287 mmio_hw_access_pre(dev_priv);
288 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
289 mmio_hw_access_post(dev_priv);
290 }
291
292 static void write_pte64(struct drm_i915_private *dev_priv,
293 unsigned long index, u64 pte)
294 {
295 void __iomem *addr = (gen8_pte_t __iomem *)dev_priv->ggtt.gsm + index;
296
297 writeq(pte, addr);
298 }
299
300 static inline int gtt_get_entry64(void *pt,
301 struct intel_gvt_gtt_entry *e,
302 unsigned long index, bool hypervisor_access, unsigned long gpa,
303 struct intel_vgpu *vgpu)
304 {
305 const struct intel_gvt_device_info *info = &vgpu->gvt->device_info;
306 int ret;
307
308 if (WARN_ON(info->gtt_entry_size != 8))
309 return -EINVAL;
310
311 if (hypervisor_access) {
312 ret = intel_gvt_hypervisor_read_gpa(vgpu, gpa +
313 (index << info->gtt_entry_size_shift),
314 &e->val64, 8);
315 if (WARN_ON(ret))
316 return ret;
317 } else if (!pt) {
318 e->val64 = read_pte64(vgpu->gvt->dev_priv, index);
319 } else {
320 e->val64 = *((u64 *)pt + index);
321 }
322 return 0;
323 }
324
325 static inline int gtt_set_entry64(void *pt,
326 struct intel_gvt_gtt_entry *e,
327 unsigned long index, bool hypervisor_access, unsigned long gpa,
328 struct intel_vgpu *vgpu)
329 {
330 const struct intel_gvt_device_info *info = &vgpu->gvt->device_info;
331 int ret;
332
333 if (WARN_ON(info->gtt_entry_size != 8))
334 return -EINVAL;
335
336 if (hypervisor_access) {
337 ret = intel_gvt_hypervisor_write_gpa(vgpu, gpa +
338 (index << info->gtt_entry_size_shift),
339 &e->val64, 8);
340 if (WARN_ON(ret))
341 return ret;
342 } else if (!pt) {
343 write_pte64(vgpu->gvt->dev_priv, index, e->val64);
344 } else {
345 *((u64 *)pt + index) = e->val64;
346 }
347 return 0;
348 }
349
350 #define GTT_HAW 46
351
352 #define ADDR_1G_MASK GENMASK_ULL(GTT_HAW - 1, 30)
353 #define ADDR_2M_MASK GENMASK_ULL(GTT_HAW - 1, 21)
354 #define ADDR_64K_MASK GENMASK_ULL(GTT_HAW - 1, 16)
355 #define ADDR_4K_MASK GENMASK_ULL(GTT_HAW - 1, 12)
356
357 #define GTT_SPTE_FLAG_MASK GENMASK_ULL(62, 52)
358 #define GTT_SPTE_FLAG_64K_SPLITED BIT(52) /* splited 64K gtt entry */
359
360 #define GTT_64K_PTE_STRIDE 16
361
362 static unsigned long gen8_gtt_get_pfn(struct intel_gvt_gtt_entry *e)
363 {
364 unsigned long pfn;
365
366 if (e->type == GTT_TYPE_PPGTT_PTE_1G_ENTRY)
367 pfn = (e->val64 & ADDR_1G_MASK) >> PAGE_SHIFT;
368 else if (e->type == GTT_TYPE_PPGTT_PTE_2M_ENTRY)
369 pfn = (e->val64 & ADDR_2M_MASK) >> PAGE_SHIFT;
370 else if (e->type == GTT_TYPE_PPGTT_PTE_64K_ENTRY)
371 pfn = (e->val64 & ADDR_64K_MASK) >> PAGE_SHIFT;
372 else
373 pfn = (e->val64 & ADDR_4K_MASK) >> PAGE_SHIFT;
374 return pfn;
375 }
376
377 static void gen8_gtt_set_pfn(struct intel_gvt_gtt_entry *e, unsigned long pfn)
378 {
379 if (e->type == GTT_TYPE_PPGTT_PTE_1G_ENTRY) {
380 e->val64 &= ~ADDR_1G_MASK;
381 pfn &= (ADDR_1G_MASK >> PAGE_SHIFT);
382 } else if (e->type == GTT_TYPE_PPGTT_PTE_2M_ENTRY) {
383 e->val64 &= ~ADDR_2M_MASK;
384 pfn &= (ADDR_2M_MASK >> PAGE_SHIFT);
385 } else if (e->type == GTT_TYPE_PPGTT_PTE_64K_ENTRY) {
386 e->val64 &= ~ADDR_64K_MASK;
387 pfn &= (ADDR_64K_MASK >> PAGE_SHIFT);
388 } else {
389 e->val64 &= ~ADDR_4K_MASK;
390 pfn &= (ADDR_4K_MASK >> PAGE_SHIFT);
391 }
392
393 e->val64 |= (pfn << PAGE_SHIFT);
394 }
395
396 static bool gen8_gtt_test_pse(struct intel_gvt_gtt_entry *e)
397 {
398 return !!(e->val64 & _PAGE_PSE);
399 }
400
401 static void gen8_gtt_clear_pse(struct intel_gvt_gtt_entry *e)
402 {
403 if (gen8_gtt_test_pse(e)) {
404 switch (e->type) {
405 case GTT_TYPE_PPGTT_PTE_2M_ENTRY:
406 e->val64 &= ~_PAGE_PSE;
407 e->type = GTT_TYPE_PPGTT_PDE_ENTRY;
408 break;
409 case GTT_TYPE_PPGTT_PTE_1G_ENTRY:
410 e->type = GTT_TYPE_PPGTT_PDP_ENTRY;
411 e->val64 &= ~_PAGE_PSE;
412 break;
413 default:
414 WARN_ON(1);
415 }
416 }
417 }
418
419 static bool gen8_gtt_test_ips(struct intel_gvt_gtt_entry *e)
420 {
421 if (GEM_WARN_ON(e->type != GTT_TYPE_PPGTT_PDE_ENTRY))
422 return false;
423
424 return !!(e->val64 & GEN8_PDE_IPS_64K);
425 }
426
427 static void gen8_gtt_clear_ips(struct intel_gvt_gtt_entry *e)
428 {
429 if (GEM_WARN_ON(e->type != GTT_TYPE_PPGTT_PDE_ENTRY))
430 return;
431
432 e->val64 &= ~GEN8_PDE_IPS_64K;
433 }
434
435 static bool gen8_gtt_test_present(struct intel_gvt_gtt_entry *e)
436 {
437 /*
438 * i915 writes PDP root pointer registers without present bit,
439 * it also works, so we need to treat root pointer entry
440 * specifically.
441 */
442 if (e->type == GTT_TYPE_PPGTT_ROOT_L3_ENTRY
443 || e->type == GTT_TYPE_PPGTT_ROOT_L4_ENTRY)
444 return (e->val64 != 0);
445 else
446 return (e->val64 & _PAGE_PRESENT);
447 }
448
449 static void gtt_entry_clear_present(struct intel_gvt_gtt_entry *e)
450 {
451 e->val64 &= ~_PAGE_PRESENT;
452 }
453
454 static void gtt_entry_set_present(struct intel_gvt_gtt_entry *e)
455 {
456 e->val64 |= _PAGE_PRESENT;
457 }
458
459 static bool gen8_gtt_test_64k_splited(struct intel_gvt_gtt_entry *e)
460 {
461 return !!(e->val64 & GTT_SPTE_FLAG_64K_SPLITED);
462 }
463
464 static void gen8_gtt_set_64k_splited(struct intel_gvt_gtt_entry *e)
465 {
466 e->val64 |= GTT_SPTE_FLAG_64K_SPLITED;
467 }
468
469 static void gen8_gtt_clear_64k_splited(struct intel_gvt_gtt_entry *e)
470 {
471 e->val64 &= ~GTT_SPTE_FLAG_64K_SPLITED;
472 }
473
474 /*
475 * Per-platform GMA routines.
476 */
477 static unsigned long gma_to_ggtt_pte_index(unsigned long gma)
478 {
479 unsigned long x = (gma >> I915_GTT_PAGE_SHIFT);
480
481 trace_gma_index(__func__, gma, x);
482 return x;
483 }
484
485 #define DEFINE_PPGTT_GMA_TO_INDEX(prefix, ename, exp) \
486 static unsigned long prefix##_gma_to_##ename##_index(unsigned long gma) \
487 { \
488 unsigned long x = (exp); \
489 trace_gma_index(__func__, gma, x); \
490 return x; \
491 }
492
493 DEFINE_PPGTT_GMA_TO_INDEX(gen8, pte, (gma >> 12 & 0x1ff));
494 DEFINE_PPGTT_GMA_TO_INDEX(gen8, pde, (gma >> 21 & 0x1ff));
495 DEFINE_PPGTT_GMA_TO_INDEX(gen8, l3_pdp, (gma >> 30 & 0x3));
496 DEFINE_PPGTT_GMA_TO_INDEX(gen8, l4_pdp, (gma >> 30 & 0x1ff));
497 DEFINE_PPGTT_GMA_TO_INDEX(gen8, pml4, (gma >> 39 & 0x1ff));
498
499 static struct intel_gvt_gtt_pte_ops gen8_gtt_pte_ops = {
500 .get_entry = gtt_get_entry64,
501 .set_entry = gtt_set_entry64,
502 .clear_present = gtt_entry_clear_present,
503 .set_present = gtt_entry_set_present,
504 .test_present = gen8_gtt_test_present,
505 .test_pse = gen8_gtt_test_pse,
506 .clear_pse = gen8_gtt_clear_pse,
507 .clear_ips = gen8_gtt_clear_ips,
508 .test_ips = gen8_gtt_test_ips,
509 .clear_64k_splited = gen8_gtt_clear_64k_splited,
510 .set_64k_splited = gen8_gtt_set_64k_splited,
511 .test_64k_splited = gen8_gtt_test_64k_splited,
512 .get_pfn = gen8_gtt_get_pfn,
513 .set_pfn = gen8_gtt_set_pfn,
514 };
515
516 static struct intel_gvt_gtt_gma_ops gen8_gtt_gma_ops = {
517 .gma_to_ggtt_pte_index = gma_to_ggtt_pte_index,
518 .gma_to_pte_index = gen8_gma_to_pte_index,
519 .gma_to_pde_index = gen8_gma_to_pde_index,
520 .gma_to_l3_pdp_index = gen8_gma_to_l3_pdp_index,
521 .gma_to_l4_pdp_index = gen8_gma_to_l4_pdp_index,
522 .gma_to_pml4_index = gen8_gma_to_pml4_index,
523 };
524
525 /* Update entry type per pse and ips bit. */
526 static void update_entry_type_for_real(struct intel_gvt_gtt_pte_ops *pte_ops,
527 struct intel_gvt_gtt_entry *entry, bool ips)
528 {
529 switch (entry->type) {
530 case GTT_TYPE_PPGTT_PDE_ENTRY:
531 case GTT_TYPE_PPGTT_PDP_ENTRY:
532 if (pte_ops->test_pse(entry))
533 entry->type = get_pse_type(entry->type);
534 break;
535 case GTT_TYPE_PPGTT_PTE_4K_ENTRY:
536 if (ips)
537 entry->type = get_pse_type(entry->type);
538 break;
539 default:
540 GEM_BUG_ON(!gtt_type_is_entry(entry->type));
541 }
542
543 GEM_BUG_ON(entry->type == GTT_TYPE_INVALID);
544 }
545
546 /*
547 * MM helpers.
548 */
549 static void _ppgtt_get_root_entry(struct intel_vgpu_mm *mm,
550 struct intel_gvt_gtt_entry *entry, unsigned long index,
551 bool guest)
552 {
553 struct intel_gvt_gtt_pte_ops *pte_ops = mm->vgpu->gvt->gtt.pte_ops;
554
555 GEM_BUG_ON(mm->type != INTEL_GVT_MM_PPGTT);
556
557 entry->type = mm->ppgtt_mm.root_entry_type;
558 pte_ops->get_entry(guest ? mm->ppgtt_mm.guest_pdps :
559 mm->ppgtt_mm.shadow_pdps,
560 entry, index, false, 0, mm->vgpu);
561 update_entry_type_for_real(pte_ops, entry, false);
562 }
563
564 static inline void ppgtt_get_guest_root_entry(struct intel_vgpu_mm *mm,
565 struct intel_gvt_gtt_entry *entry, unsigned long index)
566 {
567 _ppgtt_get_root_entry(mm, entry, index, true);
568 }
569
570 static inline void ppgtt_get_shadow_root_entry(struct intel_vgpu_mm *mm,
571 struct intel_gvt_gtt_entry *entry, unsigned long index)
572 {
573 _ppgtt_get_root_entry(mm, entry, index, false);
574 }
575
576 static void _ppgtt_set_root_entry(struct intel_vgpu_mm *mm,
577 struct intel_gvt_gtt_entry *entry, unsigned long index,
578 bool guest)
579 {
580 struct intel_gvt_gtt_pte_ops *pte_ops = mm->vgpu->gvt->gtt.pte_ops;
581
582 pte_ops->set_entry(guest ? mm->ppgtt_mm.guest_pdps :
583 mm->ppgtt_mm.shadow_pdps,
584 entry, index, false, 0, mm->vgpu);
585 }
586
587 static inline void ppgtt_set_guest_root_entry(struct intel_vgpu_mm *mm,
588 struct intel_gvt_gtt_entry *entry, unsigned long index)
589 {
590 _ppgtt_set_root_entry(mm, entry, index, true);
591 }
592
593 static inline void ppgtt_set_shadow_root_entry(struct intel_vgpu_mm *mm,
594 struct intel_gvt_gtt_entry *entry, unsigned long index)
595 {
596 _ppgtt_set_root_entry(mm, entry, index, false);
597 }
598
599 static void ggtt_get_guest_entry(struct intel_vgpu_mm *mm,
600 struct intel_gvt_gtt_entry *entry, unsigned long index)
601 {
602 struct intel_gvt_gtt_pte_ops *pte_ops = mm->vgpu->gvt->gtt.pte_ops;
603
604 GEM_BUG_ON(mm->type != INTEL_GVT_MM_GGTT);
605
606 entry->type = GTT_TYPE_GGTT_PTE;
607 pte_ops->get_entry(mm->ggtt_mm.virtual_ggtt, entry, index,
608 false, 0, mm->vgpu);
609 }
610
611 static void ggtt_set_guest_entry(struct intel_vgpu_mm *mm,
612 struct intel_gvt_gtt_entry *entry, unsigned long index)
613 {
614 struct intel_gvt_gtt_pte_ops *pte_ops = mm->vgpu->gvt->gtt.pte_ops;
615
616 GEM_BUG_ON(mm->type != INTEL_GVT_MM_GGTT);
617
618 pte_ops->set_entry(mm->ggtt_mm.virtual_ggtt, entry, index,
619 false, 0, mm->vgpu);
620 }
621
622 static void ggtt_get_host_entry(struct intel_vgpu_mm *mm,
623 struct intel_gvt_gtt_entry *entry, unsigned long index)
624 {
625 struct intel_gvt_gtt_pte_ops *pte_ops = mm->vgpu->gvt->gtt.pte_ops;
626
627 GEM_BUG_ON(mm->type != INTEL_GVT_MM_GGTT);
628
629 pte_ops->get_entry(NULL, entry, index, false, 0, mm->vgpu);
630 }
631
632 static void ggtt_set_host_entry(struct intel_vgpu_mm *mm,
633 struct intel_gvt_gtt_entry *entry, unsigned long index)
634 {
635 struct intel_gvt_gtt_pte_ops *pte_ops = mm->vgpu->gvt->gtt.pte_ops;
636
637 GEM_BUG_ON(mm->type != INTEL_GVT_MM_GGTT);
638
639 pte_ops->set_entry(NULL, entry, index, false, 0, mm->vgpu);
640 }
641
642 /*
643 * PPGTT shadow page table helpers.
644 */
645 static inline int ppgtt_spt_get_entry(
646 struct intel_vgpu_ppgtt_spt *spt,
647 void *page_table, int type,
648 struct intel_gvt_gtt_entry *e, unsigned long index,
649 bool guest)
650 {
651 struct intel_gvt *gvt = spt->vgpu->gvt;
652 struct intel_gvt_gtt_pte_ops *ops = gvt->gtt.pte_ops;
653 int ret;
654
655 e->type = get_entry_type(type);
656
657 if (WARN(!gtt_type_is_entry(e->type), "invalid entry type\n"))
658 return -EINVAL;
659
660 ret = ops->get_entry(page_table, e, index, guest,
661 spt->guest_page.gfn << I915_GTT_PAGE_SHIFT,
662 spt->vgpu);
663 if (ret)
664 return ret;
665
666 update_entry_type_for_real(ops, e, guest ?
667 spt->guest_page.pde_ips : false);
668
669 gvt_vdbg_mm("read ppgtt entry, spt type %d, entry type %d, index %lu, value %llx\n",
670 type, e->type, index, e->val64);
671 return 0;
672 }
673
674 static inline int ppgtt_spt_set_entry(
675 struct intel_vgpu_ppgtt_spt *spt,
676 void *page_table, int type,
677 struct intel_gvt_gtt_entry *e, unsigned long index,
678 bool guest)
679 {
680 struct intel_gvt *gvt = spt->vgpu->gvt;
681 struct intel_gvt_gtt_pte_ops *ops = gvt->gtt.pte_ops;
682
683 if (WARN(!gtt_type_is_entry(e->type), "invalid entry type\n"))
684 return -EINVAL;
685
686 gvt_vdbg_mm("set ppgtt entry, spt type %d, entry type %d, index %lu, value %llx\n",
687 type, e->type, index, e->val64);
688
689 return ops->set_entry(page_table, e, index, guest,
690 spt->guest_page.gfn << I915_GTT_PAGE_SHIFT,
691 spt->vgpu);
692 }
693
694 #define ppgtt_get_guest_entry(spt, e, index) \
695 ppgtt_spt_get_entry(spt, NULL, \
696 spt->guest_page.type, e, index, true)
697
698 #define ppgtt_set_guest_entry(spt, e, index) \
699 ppgtt_spt_set_entry(spt, NULL, \
700 spt->guest_page.type, e, index, true)
701
702 #define ppgtt_get_shadow_entry(spt, e, index) \
703 ppgtt_spt_get_entry(spt, spt->shadow_page.vaddr, \
704 spt->shadow_page.type, e, index, false)
705
706 #define ppgtt_set_shadow_entry(spt, e, index) \
707 ppgtt_spt_set_entry(spt, spt->shadow_page.vaddr, \
708 spt->shadow_page.type, e, index, false)
709
710 static void *alloc_spt(gfp_t gfp_mask)
711 {
712 struct intel_vgpu_ppgtt_spt *spt;
713
714 spt = kzalloc(sizeof(*spt), gfp_mask);
715 if (!spt)
716 return NULL;
717
718 spt->shadow_page.page = alloc_page(gfp_mask);
719 if (!spt->shadow_page.page) {
720 kfree(spt);
721 return NULL;
722 }
723 return spt;
724 }
725
726 static void free_spt(struct intel_vgpu_ppgtt_spt *spt)
727 {
728 __free_page(spt->shadow_page.page);
729 kfree(spt);
730 }
731
732 static int detach_oos_page(struct intel_vgpu *vgpu,
733 struct intel_vgpu_oos_page *oos_page);
734
735 static void ppgtt_free_spt(struct intel_vgpu_ppgtt_spt *spt)
736 {
737 struct device *kdev = &spt->vgpu->gvt->dev_priv->drm.pdev->dev;
738
739 trace_spt_free(spt->vgpu->id, spt, spt->guest_page.type);
740
741 dma_unmap_page(kdev, spt->shadow_page.mfn << I915_GTT_PAGE_SHIFT, 4096,
742 PCI_DMA_BIDIRECTIONAL);
743
744 radix_tree_delete(&spt->vgpu->gtt.spt_tree, spt->shadow_page.mfn);
745
746 if (spt->guest_page.gfn) {
747 if (spt->guest_page.oos_page)
748 detach_oos_page(spt->vgpu, spt->guest_page.oos_page);
749
750 intel_vgpu_unregister_page_track(spt->vgpu, spt->guest_page.gfn);
751 }
752
753 list_del_init(&spt->post_shadow_list);
754 free_spt(spt);
755 }
756
757 static void ppgtt_free_all_spt(struct intel_vgpu *vgpu)
758 {
759 struct intel_vgpu_ppgtt_spt *spt, *spn;
760 struct radix_tree_iter iter;
761 LIST_HEAD(all_spt);
762 void __rcu **slot;
763
764 rcu_read_lock();
765 radix_tree_for_each_slot(slot, &vgpu->gtt.spt_tree, &iter, 0) {
766 spt = radix_tree_deref_slot(slot);
767 list_move(&spt->post_shadow_list, &all_spt);
768 }
769 rcu_read_unlock();
770
771 list_for_each_entry_safe(spt, spn, &all_spt, post_shadow_list)
772 ppgtt_free_spt(spt);
773 }
774
775 static int ppgtt_handle_guest_write_page_table_bytes(
776 struct intel_vgpu_ppgtt_spt *spt,
777 u64 pa, void *p_data, int bytes);
778
779 static int ppgtt_write_protection_handler(
780 struct intel_vgpu_page_track *page_track,
781 u64 gpa, void *data, int bytes)
782 {
783 struct intel_vgpu_ppgtt_spt *spt = page_track->priv_data;
784
785 int ret;
786
787 if (bytes != 4 && bytes != 8)
788 return -EINVAL;
789
790 ret = ppgtt_handle_guest_write_page_table_bytes(spt, gpa, data, bytes);
791 if (ret)
792 return ret;
793 return ret;
794 }
795
796 /* Find a spt by guest gfn. */
797 static struct intel_vgpu_ppgtt_spt *intel_vgpu_find_spt_by_gfn(
798 struct intel_vgpu *vgpu, unsigned long gfn)
799 {
800 struct intel_vgpu_page_track *track;
801
802 track = intel_vgpu_find_page_track(vgpu, gfn);
803 if (track && track->handler == ppgtt_write_protection_handler)
804 return track->priv_data;
805
806 return NULL;
807 }
808
809 /* Find the spt by shadow page mfn. */
810 static inline struct intel_vgpu_ppgtt_spt *intel_vgpu_find_spt_by_mfn(
811 struct intel_vgpu *vgpu, unsigned long mfn)
812 {
813 return radix_tree_lookup(&vgpu->gtt.spt_tree, mfn);
814 }
815
816 static int reclaim_one_ppgtt_mm(struct intel_gvt *gvt);
817
818 /* Allocate shadow page table without guest page. */
819 static struct intel_vgpu_ppgtt_spt *ppgtt_alloc_spt(
820 struct intel_vgpu *vgpu, enum intel_gvt_gtt_type type)
821 {
822 struct device *kdev = &vgpu->gvt->dev_priv->drm.pdev->dev;
823 struct intel_vgpu_ppgtt_spt *spt = NULL;
824 dma_addr_t daddr;
825 int ret;
826
827 retry:
828 spt = alloc_spt(GFP_KERNEL | __GFP_ZERO);
829 if (!spt) {
830 if (reclaim_one_ppgtt_mm(vgpu->gvt))
831 goto retry;
832
833 gvt_vgpu_err("fail to allocate ppgtt shadow page\n");
834 return ERR_PTR(-ENOMEM);
835 }
836
837 spt->vgpu = vgpu;
838 atomic_set(&spt->refcount, 1);
839 INIT_LIST_HEAD(&spt->post_shadow_list);
840
841 /*
842 * Init shadow_page.
843 */
844 spt->shadow_page.type = type;
845 daddr = dma_map_page(kdev, spt->shadow_page.page,
846 0, 4096, PCI_DMA_BIDIRECTIONAL);
847 if (dma_mapping_error(kdev, daddr)) {
848 gvt_vgpu_err("fail to map dma addr\n");
849 ret = -EINVAL;
850 goto err_free_spt;
851 }
852 spt->shadow_page.vaddr = page_address(spt->shadow_page.page);
853 spt->shadow_page.mfn = daddr >> I915_GTT_PAGE_SHIFT;
854
855 ret = radix_tree_insert(&vgpu->gtt.spt_tree, spt->shadow_page.mfn, spt);
856 if (ret)
857 goto err_unmap_dma;
858
859 return spt;
860
861 err_unmap_dma:
862 dma_unmap_page(kdev, daddr, PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
863 err_free_spt:
864 free_spt(spt);
865 return ERR_PTR(ret);
866 }
867
868 /* Allocate shadow page table associated with specific gfn. */
869 static struct intel_vgpu_ppgtt_spt *ppgtt_alloc_spt_gfn(
870 struct intel_vgpu *vgpu, enum intel_gvt_gtt_type type,
871 unsigned long gfn, bool guest_pde_ips)
872 {
873 struct intel_vgpu_ppgtt_spt *spt;
874 int ret;
875
876 spt = ppgtt_alloc_spt(vgpu, type);
877 if (IS_ERR(spt))
878 return spt;
879
880 /*
881 * Init guest_page.
882 */
883 ret = intel_vgpu_register_page_track(vgpu, gfn,
884 ppgtt_write_protection_handler, spt);
885 if (ret) {
886 ppgtt_free_spt(spt);
887 return ERR_PTR(ret);
888 }
889
890 spt->guest_page.type = type;
891 spt->guest_page.gfn = gfn;
892 spt->guest_page.pde_ips = guest_pde_ips;
893
894 trace_spt_alloc(vgpu->id, spt, type, spt->shadow_page.mfn, gfn);
895
896 return spt;
897 }
898
899 #define pt_entry_size_shift(spt) \
900 ((spt)->vgpu->gvt->device_info.gtt_entry_size_shift)
901
902 #define pt_entries(spt) \
903 (I915_GTT_PAGE_SIZE >> pt_entry_size_shift(spt))
904
905 #define for_each_present_guest_entry(spt, e, i) \
906 for (i = 0; i < pt_entries(spt); \
907 i += spt->guest_page.pde_ips ? GTT_64K_PTE_STRIDE : 1) \
908 if (!ppgtt_get_guest_entry(spt, e, i) && \
909 spt->vgpu->gvt->gtt.pte_ops->test_present(e))
910
911 #define for_each_present_shadow_entry(spt, e, i) \
912 for (i = 0; i < pt_entries(spt); \
913 i += spt->shadow_page.pde_ips ? GTT_64K_PTE_STRIDE : 1) \
914 if (!ppgtt_get_shadow_entry(spt, e, i) && \
915 spt->vgpu->gvt->gtt.pte_ops->test_present(e))
916
917 #define for_each_shadow_entry(spt, e, i) \
918 for (i = 0; i < pt_entries(spt); \
919 i += (spt->shadow_page.pde_ips ? GTT_64K_PTE_STRIDE : 1)) \
920 if (!ppgtt_get_shadow_entry(spt, e, i))
921
922 static inline void ppgtt_get_spt(struct intel_vgpu_ppgtt_spt *spt)
923 {
924 int v = atomic_read(&spt->refcount);
925
926 trace_spt_refcount(spt->vgpu->id, "inc", spt, v, (v + 1));
927 atomic_inc(&spt->refcount);
928 }
929
930 static inline int ppgtt_put_spt(struct intel_vgpu_ppgtt_spt *spt)
931 {
932 int v = atomic_read(&spt->refcount);
933
934 trace_spt_refcount(spt->vgpu->id, "dec", spt, v, (v - 1));
935 return atomic_dec_return(&spt->refcount);
936 }
937
938 static int ppgtt_invalidate_spt(struct intel_vgpu_ppgtt_spt *spt);
939
940 static int ppgtt_invalidate_spt_by_shadow_entry(struct intel_vgpu *vgpu,
941 struct intel_gvt_gtt_entry *e)
942 {
943 struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
944 struct intel_vgpu_ppgtt_spt *s;
945 enum intel_gvt_gtt_type cur_pt_type;
946
947 GEM_BUG_ON(!gtt_type_is_pt(get_next_pt_type(e->type)));
948
949 if (e->type != GTT_TYPE_PPGTT_ROOT_L3_ENTRY
950 && e->type != GTT_TYPE_PPGTT_ROOT_L4_ENTRY) {
951 cur_pt_type = get_next_pt_type(e->type);
952
953 if (!gtt_type_is_pt(cur_pt_type) ||
954 !gtt_type_is_pt(cur_pt_type + 1)) {
955 WARN(1, "Invalid page table type, cur_pt_type is: %d\n", cur_pt_type);
956 return -EINVAL;
957 }
958
959 cur_pt_type += 1;
960
961 if (ops->get_pfn(e) ==
962 vgpu->gtt.scratch_pt[cur_pt_type].page_mfn)
963 return 0;
964 }
965 s = intel_vgpu_find_spt_by_mfn(vgpu, ops->get_pfn(e));
966 if (!s) {
967 gvt_vgpu_err("fail to find shadow page: mfn: 0x%lx\n",
968 ops->get_pfn(e));
969 return -ENXIO;
970 }
971 return ppgtt_invalidate_spt(s);
972 }
973
974 static inline void ppgtt_invalidate_pte(struct intel_vgpu_ppgtt_spt *spt,
975 struct intel_gvt_gtt_entry *entry)
976 {
977 struct intel_vgpu *vgpu = spt->vgpu;
978 struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
979 unsigned long pfn;
980 int type;
981
982 pfn = ops->get_pfn(entry);
983 type = spt->shadow_page.type;
984
985 /* Uninitialized spte or unshadowed spte. */
986 if (!pfn || pfn == vgpu->gtt.scratch_pt[type].page_mfn)
987 return;
988
989 intel_gvt_hypervisor_dma_unmap_guest_page(vgpu, pfn << PAGE_SHIFT);
990 }
991
992 static int ppgtt_invalidate_spt(struct intel_vgpu_ppgtt_spt *spt)
993 {
994 struct intel_vgpu *vgpu = spt->vgpu;
995 struct intel_gvt_gtt_entry e;
996 unsigned long index;
997 int ret;
998
999 trace_spt_change(spt->vgpu->id, "die", spt,
1000 spt->guest_page.gfn, spt->shadow_page.type);
1001
1002 if (ppgtt_put_spt(spt) > 0)
1003 return 0;
1004
1005 for_each_present_shadow_entry(spt, &e, index) {
1006 switch (e.type) {
1007 case GTT_TYPE_PPGTT_PTE_4K_ENTRY:
1008 gvt_vdbg_mm("invalidate 4K entry\n");
1009 ppgtt_invalidate_pte(spt, &e);
1010 break;
1011 case GTT_TYPE_PPGTT_PTE_64K_ENTRY:
1012 /* We don't setup 64K shadow entry so far. */
1013 WARN(1, "suspicious 64K gtt entry\n");
1014 continue;
1015 case GTT_TYPE_PPGTT_PTE_2M_ENTRY:
1016 gvt_vdbg_mm("invalidate 2M entry\n");
1017 continue;
1018 case GTT_TYPE_PPGTT_PTE_1G_ENTRY:
1019 WARN(1, "GVT doesn't support 1GB page\n");
1020 continue;
1021 case GTT_TYPE_PPGTT_PML4_ENTRY:
1022 case GTT_TYPE_PPGTT_PDP_ENTRY:
1023 case GTT_TYPE_PPGTT_PDE_ENTRY:
1024 gvt_vdbg_mm("invalidate PMUL4/PDP/PDE entry\n");
1025 ret = ppgtt_invalidate_spt_by_shadow_entry(
1026 spt->vgpu, &e);
1027 if (ret)
1028 goto fail;
1029 break;
1030 default:
1031 GEM_BUG_ON(1);
1032 }
1033 }
1034
1035 trace_spt_change(spt->vgpu->id, "release", spt,
1036 spt->guest_page.gfn, spt->shadow_page.type);
1037 ppgtt_free_spt(spt);
1038 return 0;
1039 fail:
1040 gvt_vgpu_err("fail: shadow page %p shadow entry 0x%llx type %d\n",
1041 spt, e.val64, e.type);
1042 return ret;
1043 }
1044
1045 static bool vgpu_ips_enabled(struct intel_vgpu *vgpu)
1046 {
1047 struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
1048
1049 if (INTEL_GEN(dev_priv) == 9 || INTEL_GEN(dev_priv) == 10) {
1050 u32 ips = vgpu_vreg_t(vgpu, GEN8_GAMW_ECO_DEV_RW_IA) &
1051 GAMW_ECO_ENABLE_64K_IPS_FIELD;
1052
1053 return ips == GAMW_ECO_ENABLE_64K_IPS_FIELD;
1054 } else if (INTEL_GEN(dev_priv) >= 11) {
1055 /* 64K paging only controlled by IPS bit in PTE now. */
1056 return true;
1057 } else
1058 return false;
1059 }
1060
1061 static int ppgtt_populate_spt(struct intel_vgpu_ppgtt_spt *spt);
1062
1063 static struct intel_vgpu_ppgtt_spt *ppgtt_populate_spt_by_guest_entry(
1064 struct intel_vgpu *vgpu, struct intel_gvt_gtt_entry *we)
1065 {
1066 struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
1067 struct intel_vgpu_ppgtt_spt *spt = NULL;
1068 bool ips = false;
1069 int ret;
1070
1071 GEM_BUG_ON(!gtt_type_is_pt(get_next_pt_type(we->type)));
1072
1073 if (we->type == GTT_TYPE_PPGTT_PDE_ENTRY)
1074 ips = vgpu_ips_enabled(vgpu) && ops->test_ips(we);
1075
1076 spt = intel_vgpu_find_spt_by_gfn(vgpu, ops->get_pfn(we));
1077 if (spt) {
1078 ppgtt_get_spt(spt);
1079
1080 if (ips != spt->guest_page.pde_ips) {
1081 spt->guest_page.pde_ips = ips;
1082
1083 gvt_dbg_mm("reshadow PDE since ips changed\n");
1084 clear_page(spt->shadow_page.vaddr);
1085 ret = ppgtt_populate_spt(spt);
1086 if (ret) {
1087 ppgtt_put_spt(spt);
1088 goto err;
1089 }
1090 }
1091 } else {
1092 int type = get_next_pt_type(we->type);
1093
1094 if (!gtt_type_is_pt(type)) {
1095 ret = -EINVAL;
1096 goto err;
1097 }
1098
1099 spt = ppgtt_alloc_spt_gfn(vgpu, type, ops->get_pfn(we), ips);
1100 if (IS_ERR(spt)) {
1101 ret = PTR_ERR(spt);
1102 goto err;
1103 }
1104
1105 ret = intel_vgpu_enable_page_track(vgpu, spt->guest_page.gfn);
1106 if (ret)
1107 goto err_free_spt;
1108
1109 ret = ppgtt_populate_spt(spt);
1110 if (ret)
1111 goto err_free_spt;
1112
1113 trace_spt_change(vgpu->id, "new", spt, spt->guest_page.gfn,
1114 spt->shadow_page.type);
1115 }
1116 return spt;
1117
1118 err_free_spt:
1119 ppgtt_free_spt(spt);
1120 spt = NULL;
1121 err:
1122 gvt_vgpu_err("fail: shadow page %p guest entry 0x%llx type %d\n",
1123 spt, we->val64, we->type);
1124 return ERR_PTR(ret);
1125 }
1126
1127 static inline void ppgtt_generate_shadow_entry(struct intel_gvt_gtt_entry *se,
1128 struct intel_vgpu_ppgtt_spt *s, struct intel_gvt_gtt_entry *ge)
1129 {
1130 struct intel_gvt_gtt_pte_ops *ops = s->vgpu->gvt->gtt.pte_ops;
1131
1132 se->type = ge->type;
1133 se->val64 = ge->val64;
1134
1135 /* Because we always split 64KB pages, so clear IPS in shadow PDE. */
1136 if (se->type == GTT_TYPE_PPGTT_PDE_ENTRY)
1137 ops->clear_ips(se);
1138
1139 ops->set_pfn(se, s->shadow_page.mfn);
1140 }
1141
1142 /**
1143 * Check if can do 2M page
1144 * @vgpu: target vgpu
1145 * @entry: target pfn's gtt entry
1146 *
1147 * Return 1 if 2MB huge gtt shadowing is possilbe, 0 if miscondition,
1148 * negtive if found err.
1149 */
1150 static int is_2MB_gtt_possible(struct intel_vgpu *vgpu,
1151 struct intel_gvt_gtt_entry *entry)
1152 {
1153 struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
1154 unsigned long pfn;
1155
1156 if (!HAS_PAGE_SIZES(vgpu->gvt->dev_priv, I915_GTT_PAGE_SIZE_2M))
1157 return 0;
1158
1159 pfn = intel_gvt_hypervisor_gfn_to_mfn(vgpu, ops->get_pfn(entry));
1160 if (pfn == INTEL_GVT_INVALID_ADDR)
1161 return -EINVAL;
1162
1163 return PageTransHuge(pfn_to_page(pfn));
1164 }
1165
1166 static int split_2MB_gtt_entry(struct intel_vgpu *vgpu,
1167 struct intel_vgpu_ppgtt_spt *spt, unsigned long index,
1168 struct intel_gvt_gtt_entry *se)
1169 {
1170 struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
1171 struct intel_vgpu_ppgtt_spt *sub_spt;
1172 struct intel_gvt_gtt_entry sub_se;
1173 unsigned long start_gfn;
1174 dma_addr_t dma_addr;
1175 unsigned long sub_index;
1176 int ret;
1177
1178 gvt_dbg_mm("Split 2M gtt entry, index %lu\n", index);
1179
1180 start_gfn = ops->get_pfn(se);
1181
1182 sub_spt = ppgtt_alloc_spt(vgpu, GTT_TYPE_PPGTT_PTE_PT);
1183 if (IS_ERR(sub_spt))
1184 return PTR_ERR(sub_spt);
1185
1186 for_each_shadow_entry(sub_spt, &sub_se, sub_index) {
1187 ret = intel_gvt_hypervisor_dma_map_guest_page(vgpu,
1188 start_gfn + sub_index, PAGE_SIZE, &dma_addr);
1189 if (ret) {
1190 ppgtt_invalidate_spt(spt);
1191 return ret;
1192 }
1193 sub_se.val64 = se->val64;
1194
1195 /* Copy the PAT field from PDE. */
1196 sub_se.val64 &= ~_PAGE_PAT;
1197 sub_se.val64 |= (se->val64 & _PAGE_PAT_LARGE) >> 5;
1198
1199 ops->set_pfn(&sub_se, dma_addr >> PAGE_SHIFT);
1200 ppgtt_set_shadow_entry(sub_spt, &sub_se, sub_index);
1201 }
1202
1203 /* Clear dirty field. */
1204 se->val64 &= ~_PAGE_DIRTY;
1205
1206 ops->clear_pse(se);
1207 ops->clear_ips(se);
1208 ops->set_pfn(se, sub_spt->shadow_page.mfn);
1209 ppgtt_set_shadow_entry(spt, se, index);
1210 return 0;
1211 }
1212
1213 static int split_64KB_gtt_entry(struct intel_vgpu *vgpu,
1214 struct intel_vgpu_ppgtt_spt *spt, unsigned long index,
1215 struct intel_gvt_gtt_entry *se)
1216 {
1217 struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
1218 struct intel_gvt_gtt_entry entry = *se;
1219 unsigned long start_gfn;
1220 dma_addr_t dma_addr;
1221 int i, ret;
1222
1223 gvt_vdbg_mm("Split 64K gtt entry, index %lu\n", index);
1224
1225 GEM_BUG_ON(index % GTT_64K_PTE_STRIDE);
1226
1227 start_gfn = ops->get_pfn(se);
1228
1229 entry.type = GTT_TYPE_PPGTT_PTE_4K_ENTRY;
1230 ops->set_64k_splited(&entry);
1231
1232 for (i = 0; i < GTT_64K_PTE_STRIDE; i++) {
1233 ret = intel_gvt_hypervisor_dma_map_guest_page(vgpu,
1234 start_gfn + i, PAGE_SIZE, &dma_addr);
1235 if (ret)
1236 return ret;
1237
1238 ops->set_pfn(&entry, dma_addr >> PAGE_SHIFT);
1239 ppgtt_set_shadow_entry(spt, &entry, index + i);
1240 }
1241 return 0;
1242 }
1243
1244 static int ppgtt_populate_shadow_entry(struct intel_vgpu *vgpu,
1245 struct intel_vgpu_ppgtt_spt *spt, unsigned long index,
1246 struct intel_gvt_gtt_entry *ge)
1247 {
1248 struct intel_gvt_gtt_pte_ops *pte_ops = vgpu->gvt->gtt.pte_ops;
1249 struct intel_gvt_gtt_entry se = *ge;
1250 unsigned long gfn, page_size = PAGE_SIZE;
1251 dma_addr_t dma_addr;
1252 int ret;
1253
1254 if (!pte_ops->test_present(ge))
1255 return 0;
1256
1257 gfn = pte_ops->get_pfn(ge);
1258
1259 switch (ge->type) {
1260 case GTT_TYPE_PPGTT_PTE_4K_ENTRY:
1261 gvt_vdbg_mm("shadow 4K gtt entry\n");
1262 break;
1263 case GTT_TYPE_PPGTT_PTE_64K_ENTRY:
1264 gvt_vdbg_mm("shadow 64K gtt entry\n");
1265 /*
1266 * The layout of 64K page is special, the page size is
1267 * controlled by uper PDE. To be simple, we always split
1268 * 64K page to smaller 4K pages in shadow PT.
1269 */
1270 return split_64KB_gtt_entry(vgpu, spt, index, &se);
1271 case GTT_TYPE_PPGTT_PTE_2M_ENTRY:
1272 gvt_vdbg_mm("shadow 2M gtt entry\n");
1273 ret = is_2MB_gtt_possible(vgpu, ge);
1274 if (ret == 0)
1275 return split_2MB_gtt_entry(vgpu, spt, index, &se);
1276 else if (ret < 0)
1277 return ret;
1278 page_size = I915_GTT_PAGE_SIZE_2M;
1279 break;
1280 case GTT_TYPE_PPGTT_PTE_1G_ENTRY:
1281 gvt_vgpu_err("GVT doesn't support 1GB entry\n");
1282 return -EINVAL;
1283 default:
1284 GEM_BUG_ON(1);
1285 }
1286
1287 /* direct shadow */
1288 ret = intel_gvt_hypervisor_dma_map_guest_page(vgpu, gfn, page_size,
1289 &dma_addr);
1290 if (ret)
1291 return -ENXIO;
1292
1293 pte_ops->set_pfn(&se, dma_addr >> PAGE_SHIFT);
1294 ppgtt_set_shadow_entry(spt, &se, index);
1295 return 0;
1296 }
1297
1298 static int ppgtt_populate_spt(struct intel_vgpu_ppgtt_spt *spt)
1299 {
1300 struct intel_vgpu *vgpu = spt->vgpu;
1301 struct intel_gvt *gvt = vgpu->gvt;
1302 struct intel_gvt_gtt_pte_ops *ops = gvt->gtt.pte_ops;
1303 struct intel_vgpu_ppgtt_spt *s;
1304 struct intel_gvt_gtt_entry se, ge;
1305 unsigned long gfn, i;
1306 int ret;
1307
1308 trace_spt_change(spt->vgpu->id, "born", spt,
1309 spt->guest_page.gfn, spt->shadow_page.type);
1310
1311 for_each_present_guest_entry(spt, &ge, i) {
1312 if (gtt_type_is_pt(get_next_pt_type(ge.type))) {
1313 s = ppgtt_populate_spt_by_guest_entry(vgpu, &ge);
1314 if (IS_ERR(s)) {
1315 ret = PTR_ERR(s);
1316 goto fail;
1317 }
1318 ppgtt_get_shadow_entry(spt, &se, i);
1319 ppgtt_generate_shadow_entry(&se, s, &ge);
1320 ppgtt_set_shadow_entry(spt, &se, i);
1321 } else {
1322 gfn = ops->get_pfn(&ge);
1323 if (!intel_gvt_hypervisor_is_valid_gfn(vgpu, gfn)) {
1324 ops->set_pfn(&se, gvt->gtt.scratch_mfn);
1325 ppgtt_set_shadow_entry(spt, &se, i);
1326 continue;
1327 }
1328
1329 ret = ppgtt_populate_shadow_entry(vgpu, spt, i, &ge);
1330 if (ret)
1331 goto fail;
1332 }
1333 }
1334 return 0;
1335 fail:
1336 gvt_vgpu_err("fail: shadow page %p guest entry 0x%llx type %d\n",
1337 spt, ge.val64, ge.type);
1338 return ret;
1339 }
1340
1341 static int ppgtt_handle_guest_entry_removal(struct intel_vgpu_ppgtt_spt *spt,
1342 struct intel_gvt_gtt_entry *se, unsigned long index)
1343 {
1344 struct intel_vgpu *vgpu = spt->vgpu;
1345 struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
1346 int ret;
1347
1348 trace_spt_guest_change(spt->vgpu->id, "remove", spt,
1349 spt->shadow_page.type, se->val64, index);
1350
1351 gvt_vdbg_mm("destroy old shadow entry, type %d, index %lu, value %llx\n",
1352 se->type, index, se->val64);
1353
1354 if (!ops->test_present(se))
1355 return 0;
1356
1357 if (ops->get_pfn(se) ==
1358 vgpu->gtt.scratch_pt[spt->shadow_page.type].page_mfn)
1359 return 0;
1360
1361 if (gtt_type_is_pt(get_next_pt_type(se->type))) {
1362 struct intel_vgpu_ppgtt_spt *s =
1363 intel_vgpu_find_spt_by_mfn(vgpu, ops->get_pfn(se));
1364 if (!s) {
1365 gvt_vgpu_err("fail to find guest page\n");
1366 ret = -ENXIO;
1367 goto fail;
1368 }
1369 ret = ppgtt_invalidate_spt(s);
1370 if (ret)
1371 goto fail;
1372 } else {
1373 /* We don't setup 64K shadow entry so far. */
1374 WARN(se->type == GTT_TYPE_PPGTT_PTE_64K_ENTRY,
1375 "suspicious 64K entry\n");
1376 ppgtt_invalidate_pte(spt, se);
1377 }
1378
1379 return 0;
1380 fail:
1381 gvt_vgpu_err("fail: shadow page %p guest entry 0x%llx type %d\n",
1382 spt, se->val64, se->type);
1383 return ret;
1384 }
1385
1386 static int ppgtt_handle_guest_entry_add(struct intel_vgpu_ppgtt_spt *spt,
1387 struct intel_gvt_gtt_entry *we, unsigned long index)
1388 {
1389 struct intel_vgpu *vgpu = spt->vgpu;
1390 struct intel_gvt_gtt_entry m;
1391 struct intel_vgpu_ppgtt_spt *s;
1392 int ret;
1393
1394 trace_spt_guest_change(spt->vgpu->id, "add", spt, spt->shadow_page.type,
1395 we->val64, index);
1396
1397 gvt_vdbg_mm("add shadow entry: type %d, index %lu, value %llx\n",
1398 we->type, index, we->val64);
1399
1400 if (gtt_type_is_pt(get_next_pt_type(we->type))) {
1401 s = ppgtt_populate_spt_by_guest_entry(vgpu, we);
1402 if (IS_ERR(s)) {
1403 ret = PTR_ERR(s);
1404 goto fail;
1405 }
1406 ppgtt_get_shadow_entry(spt, &m, index);
1407 ppgtt_generate_shadow_entry(&m, s, we);
1408 ppgtt_set_shadow_entry(spt, &m, index);
1409 } else {
1410 ret = ppgtt_populate_shadow_entry(vgpu, spt, index, we);
1411 if (ret)
1412 goto fail;
1413 }
1414 return 0;
1415 fail:
1416 gvt_vgpu_err("fail: spt %p guest entry 0x%llx type %d\n",
1417 spt, we->val64, we->type);
1418 return ret;
1419 }
1420
1421 static int sync_oos_page(struct intel_vgpu *vgpu,
1422 struct intel_vgpu_oos_page *oos_page)
1423 {
1424 const struct intel_gvt_device_info *info = &vgpu->gvt->device_info;
1425 struct intel_gvt *gvt = vgpu->gvt;
1426 struct intel_gvt_gtt_pte_ops *ops = gvt->gtt.pte_ops;
1427 struct intel_vgpu_ppgtt_spt *spt = oos_page->spt;
1428 struct intel_gvt_gtt_entry old, new;
1429 int index;
1430 int ret;
1431
1432 trace_oos_change(vgpu->id, "sync", oos_page->id,
1433 spt, spt->guest_page.type);
1434
1435 old.type = new.type = get_entry_type(spt->guest_page.type);
1436 old.val64 = new.val64 = 0;
1437
1438 for (index = 0; index < (I915_GTT_PAGE_SIZE >>
1439 info->gtt_entry_size_shift); index++) {
1440 ops->get_entry(oos_page->mem, &old, index, false, 0, vgpu);
1441 ops->get_entry(NULL, &new, index, true,
1442 spt->guest_page.gfn << PAGE_SHIFT, vgpu);
1443
1444 if (old.val64 == new.val64
1445 && !test_and_clear_bit(index, spt->post_shadow_bitmap))
1446 continue;
1447
1448 trace_oos_sync(vgpu->id, oos_page->id,
1449 spt, spt->guest_page.type,
1450 new.val64, index);
1451
1452 ret = ppgtt_populate_shadow_entry(vgpu, spt, index, &new);
1453 if (ret)
1454 return ret;
1455
1456 ops->set_entry(oos_page->mem, &new, index, false, 0, vgpu);
1457 }
1458
1459 spt->guest_page.write_cnt = 0;
1460 list_del_init(&spt->post_shadow_list);
1461 return 0;
1462 }
1463
1464 static int detach_oos_page(struct intel_vgpu *vgpu,
1465 struct intel_vgpu_oos_page *oos_page)
1466 {
1467 struct intel_gvt *gvt = vgpu->gvt;
1468 struct intel_vgpu_ppgtt_spt *spt = oos_page->spt;
1469
1470 trace_oos_change(vgpu->id, "detach", oos_page->id,
1471 spt, spt->guest_page.type);
1472
1473 spt->guest_page.write_cnt = 0;
1474 spt->guest_page.oos_page = NULL;
1475 oos_page->spt = NULL;
1476
1477 list_del_init(&oos_page->vm_list);
1478 list_move_tail(&oos_page->list, &gvt->gtt.oos_page_free_list_head);
1479
1480 return 0;
1481 }
1482
1483 static int attach_oos_page(struct intel_vgpu_oos_page *oos_page,
1484 struct intel_vgpu_ppgtt_spt *spt)
1485 {
1486 struct intel_gvt *gvt = spt->vgpu->gvt;
1487 int ret;
1488
1489 ret = intel_gvt_hypervisor_read_gpa(spt->vgpu,
1490 spt->guest_page.gfn << I915_GTT_PAGE_SHIFT,
1491 oos_page->mem, I915_GTT_PAGE_SIZE);
1492 if (ret)
1493 return ret;
1494
1495 oos_page->spt = spt;
1496 spt->guest_page.oos_page = oos_page;
1497
1498 list_move_tail(&oos_page->list, &gvt->gtt.oos_page_use_list_head);
1499
1500 trace_oos_change(spt->vgpu->id, "attach", oos_page->id,
1501 spt, spt->guest_page.type);
1502 return 0;
1503 }
1504
1505 static int ppgtt_set_guest_page_sync(struct intel_vgpu_ppgtt_spt *spt)
1506 {
1507 struct intel_vgpu_oos_page *oos_page = spt->guest_page.oos_page;
1508 int ret;
1509
1510 ret = intel_vgpu_enable_page_track(spt->vgpu, spt->guest_page.gfn);
1511 if (ret)
1512 return ret;
1513
1514 trace_oos_change(spt->vgpu->id, "set page sync", oos_page->id,
1515 spt, spt->guest_page.type);
1516
1517 list_del_init(&oos_page->vm_list);
1518 return sync_oos_page(spt->vgpu, oos_page);
1519 }
1520
1521 static int ppgtt_allocate_oos_page(struct intel_vgpu_ppgtt_spt *spt)
1522 {
1523 struct intel_gvt *gvt = spt->vgpu->gvt;
1524 struct intel_gvt_gtt *gtt = &gvt->gtt;
1525 struct intel_vgpu_oos_page *oos_page = spt->guest_page.oos_page;
1526 int ret;
1527
1528 WARN(oos_page, "shadow PPGTT page has already has a oos page\n");
1529
1530 if (list_empty(&gtt->oos_page_free_list_head)) {
1531 oos_page = container_of(gtt->oos_page_use_list_head.next,
1532 struct intel_vgpu_oos_page, list);
1533 ret = ppgtt_set_guest_page_sync(oos_page->spt);
1534 if (ret)
1535 return ret;
1536 ret = detach_oos_page(spt->vgpu, oos_page);
1537 if (ret)
1538 return ret;
1539 } else
1540 oos_page = container_of(gtt->oos_page_free_list_head.next,
1541 struct intel_vgpu_oos_page, list);
1542 return attach_oos_page(oos_page, spt);
1543 }
1544
1545 static int ppgtt_set_guest_page_oos(struct intel_vgpu_ppgtt_spt *spt)
1546 {
1547 struct intel_vgpu_oos_page *oos_page = spt->guest_page.oos_page;
1548
1549 if (WARN(!oos_page, "shadow PPGTT page should have a oos page\n"))
1550 return -EINVAL;
1551
1552 trace_oos_change(spt->vgpu->id, "set page out of sync", oos_page->id,
1553 spt, spt->guest_page.type);
1554
1555 list_add_tail(&oos_page->vm_list, &spt->vgpu->gtt.oos_page_list_head);
1556 return intel_vgpu_disable_page_track(spt->vgpu, spt->guest_page.gfn);
1557 }
1558
1559 /**
1560 * intel_vgpu_sync_oos_pages - sync all the out-of-synced shadow for vGPU
1561 * @vgpu: a vGPU
1562 *
1563 * This function is called before submitting a guest workload to host,
1564 * to sync all the out-of-synced shadow for vGPU
1565 *
1566 * Returns:
1567 * Zero on success, negative error code if failed.
1568 */
1569 int intel_vgpu_sync_oos_pages(struct intel_vgpu *vgpu)
1570 {
1571 struct list_head *pos, *n;
1572 struct intel_vgpu_oos_page *oos_page;
1573 int ret;
1574
1575 if (!enable_out_of_sync)
1576 return 0;
1577
1578 list_for_each_safe(pos, n, &vgpu->gtt.oos_page_list_head) {
1579 oos_page = container_of(pos,
1580 struct intel_vgpu_oos_page, vm_list);
1581 ret = ppgtt_set_guest_page_sync(oos_page->spt);
1582 if (ret)
1583 return ret;
1584 }
1585 return 0;
1586 }
1587
1588 /*
1589 * The heart of PPGTT shadow page table.
1590 */
1591 static int ppgtt_handle_guest_write_page_table(
1592 struct intel_vgpu_ppgtt_spt *spt,
1593 struct intel_gvt_gtt_entry *we, unsigned long index)
1594 {
1595 struct intel_vgpu *vgpu = spt->vgpu;
1596 int type = spt->shadow_page.type;
1597 struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
1598 struct intel_gvt_gtt_entry old_se;
1599 int new_present;
1600 int i, ret;
1601
1602 new_present = ops->test_present(we);
1603
1604 /*
1605 * Adding the new entry first and then removing the old one, that can
1606 * guarantee the ppgtt table is validated during the window between
1607 * adding and removal.
1608 */
1609 ppgtt_get_shadow_entry(spt, &old_se, index);
1610
1611 if (new_present) {
1612 ret = ppgtt_handle_guest_entry_add(spt, we, index);
1613 if (ret)
1614 goto fail;
1615 }
1616
1617 ret = ppgtt_handle_guest_entry_removal(spt, &old_se, index);
1618 if (ret)
1619 goto fail;
1620
1621 if (!new_present) {
1622 /* For 64KB splited entries, we need clear them all. */
1623 if (ops->test_64k_splited(&old_se) &&
1624 !(index % GTT_64K_PTE_STRIDE)) {
1625 gvt_vdbg_mm("remove splited 64K shadow entries\n");
1626 for (i = 0; i < GTT_64K_PTE_STRIDE; i++) {
1627 ops->clear_64k_splited(&old_se);
1628 ops->set_pfn(&old_se,
1629 vgpu->gtt.scratch_pt[type].page_mfn);
1630 ppgtt_set_shadow_entry(spt, &old_se, index + i);
1631 }
1632 } else if (old_se.type == GTT_TYPE_PPGTT_PTE_2M_ENTRY ||
1633 old_se.type == GTT_TYPE_PPGTT_PTE_1G_ENTRY) {
1634 ops->clear_pse(&old_se);
1635 ops->set_pfn(&old_se,
1636 vgpu->gtt.scratch_pt[type].page_mfn);
1637 ppgtt_set_shadow_entry(spt, &old_se, index);
1638 } else {
1639 ops->set_pfn(&old_se,
1640 vgpu->gtt.scratch_pt[type].page_mfn);
1641 ppgtt_set_shadow_entry(spt, &old_se, index);
1642 }
1643 }
1644
1645 return 0;
1646 fail:
1647 gvt_vgpu_err("fail: shadow page %p guest entry 0x%llx type %d.\n",
1648 spt, we->val64, we->type);
1649 return ret;
1650 }
1651
1652
1653
1654 static inline bool can_do_out_of_sync(struct intel_vgpu_ppgtt_spt *spt)
1655 {
1656 return enable_out_of_sync
1657 && gtt_type_is_pte_pt(spt->guest_page.type)
1658 && spt->guest_page.write_cnt >= 2;
1659 }
1660
1661 static void ppgtt_set_post_shadow(struct intel_vgpu_ppgtt_spt *spt,
1662 unsigned long index)
1663 {
1664 set_bit(index, spt->post_shadow_bitmap);
1665 if (!list_empty(&spt->post_shadow_list))
1666 return;
1667
1668 list_add_tail(&spt->post_shadow_list,
1669 &spt->vgpu->gtt.post_shadow_list_head);
1670 }
1671
1672 /**
1673 * intel_vgpu_flush_post_shadow - flush the post shadow transactions
1674 * @vgpu: a vGPU
1675 *
1676 * This function is called before submitting a guest workload to host,
1677 * to flush all the post shadows for a vGPU.
1678 *
1679 * Returns:
1680 * Zero on success, negative error code if failed.
1681 */
1682 int intel_vgpu_flush_post_shadow(struct intel_vgpu *vgpu)
1683 {
1684 struct list_head *pos, *n;
1685 struct intel_vgpu_ppgtt_spt *spt;
1686 struct intel_gvt_gtt_entry ge;
1687 unsigned long index;
1688 int ret;
1689
1690 list_for_each_safe(pos, n, &vgpu->gtt.post_shadow_list_head) {
1691 spt = container_of(pos, struct intel_vgpu_ppgtt_spt,
1692 post_shadow_list);
1693
1694 for_each_set_bit(index, spt->post_shadow_bitmap,
1695 GTT_ENTRY_NUM_IN_ONE_PAGE) {
1696 ppgtt_get_guest_entry(spt, &ge, index);
1697
1698 ret = ppgtt_handle_guest_write_page_table(spt,
1699 &ge, index);
1700 if (ret)
1701 return ret;
1702 clear_bit(index, spt->post_shadow_bitmap);
1703 }
1704 list_del_init(&spt->post_shadow_list);
1705 }
1706 return 0;
1707 }
1708
1709 static int ppgtt_handle_guest_write_page_table_bytes(
1710 struct intel_vgpu_ppgtt_spt *spt,
1711 u64 pa, void *p_data, int bytes)
1712 {
1713 struct intel_vgpu *vgpu = spt->vgpu;
1714 struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
1715 const struct intel_gvt_device_info *info = &vgpu->gvt->device_info;
1716 struct intel_gvt_gtt_entry we, se;
1717 unsigned long index;
1718 int ret;
1719
1720 index = (pa & (PAGE_SIZE - 1)) >> info->gtt_entry_size_shift;
1721
1722 ppgtt_get_guest_entry(spt, &we, index);
1723
1724 /*
1725 * For page table which has 64K gtt entry, only PTE#0, PTE#16,
1726 * PTE#32, ... PTE#496 are used. Unused PTEs update should be
1727 * ignored.
1728 */
1729 if (we.type == GTT_TYPE_PPGTT_PTE_64K_ENTRY &&
1730 (index % GTT_64K_PTE_STRIDE)) {
1731 gvt_vdbg_mm("Ignore write to unused PTE entry, index %lu\n",
1732 index);
1733 return 0;
1734 }
1735
1736 if (bytes == info->gtt_entry_size) {
1737 ret = ppgtt_handle_guest_write_page_table(spt, &we, index);
1738 if (ret)
1739 return ret;
1740 } else {
1741 if (!test_bit(index, spt->post_shadow_bitmap)) {
1742 int type = spt->shadow_page.type;
1743
1744 ppgtt_get_shadow_entry(spt, &se, index);
1745 ret = ppgtt_handle_guest_entry_removal(spt, &se, index);
1746 if (ret)
1747 return ret;
1748 ops->set_pfn(&se, vgpu->gtt.scratch_pt[type].page_mfn);
1749 ppgtt_set_shadow_entry(spt, &se, index);
1750 }
1751 ppgtt_set_post_shadow(spt, index);
1752 }
1753
1754 if (!enable_out_of_sync)
1755 return 0;
1756
1757 spt->guest_page.write_cnt++;
1758
1759 if (spt->guest_page.oos_page)
1760 ops->set_entry(spt->guest_page.oos_page->mem, &we, index,
1761 false, 0, vgpu);
1762
1763 if (can_do_out_of_sync(spt)) {
1764 if (!spt->guest_page.oos_page)
1765 ppgtt_allocate_oos_page(spt);
1766
1767 ret = ppgtt_set_guest_page_oos(spt);
1768 if (ret < 0)
1769 return ret;
1770 }
1771 return 0;
1772 }
1773
1774 static void invalidate_ppgtt_mm(struct intel_vgpu_mm *mm)
1775 {
1776 struct intel_vgpu *vgpu = mm->vgpu;
1777 struct intel_gvt *gvt = vgpu->gvt;
1778 struct intel_gvt_gtt *gtt = &gvt->gtt;
1779 struct intel_gvt_gtt_pte_ops *ops = gtt->pte_ops;
1780 struct intel_gvt_gtt_entry se;
1781 int index;
1782
1783 if (!mm->ppgtt_mm.shadowed)
1784 return;
1785
1786 for (index = 0; index < ARRAY_SIZE(mm->ppgtt_mm.shadow_pdps); index++) {
1787 ppgtt_get_shadow_root_entry(mm, &se, index);
1788
1789 if (!ops->test_present(&se))
1790 continue;
1791
1792 ppgtt_invalidate_spt_by_shadow_entry(vgpu, &se);
1793 se.val64 = 0;
1794 ppgtt_set_shadow_root_entry(mm, &se, index);
1795
1796 trace_spt_guest_change(vgpu->id, "destroy root pointer",
1797 NULL, se.type, se.val64, index);
1798 }
1799
1800 mm->ppgtt_mm.shadowed = false;
1801 }
1802
1803
1804 static int shadow_ppgtt_mm(struct intel_vgpu_mm *mm)
1805 {
1806 struct intel_vgpu *vgpu = mm->vgpu;
1807 struct intel_gvt *gvt = vgpu->gvt;
1808 struct intel_gvt_gtt *gtt = &gvt->gtt;
1809 struct intel_gvt_gtt_pte_ops *ops = gtt->pte_ops;
1810 struct intel_vgpu_ppgtt_spt *spt;
1811 struct intel_gvt_gtt_entry ge, se;
1812 int index, ret;
1813
1814 if (mm->ppgtt_mm.shadowed)
1815 return 0;
1816
1817 mm->ppgtt_mm.shadowed = true;
1818
1819 for (index = 0; index < ARRAY_SIZE(mm->ppgtt_mm.guest_pdps); index++) {
1820 ppgtt_get_guest_root_entry(mm, &ge, index);
1821
1822 if (!ops->test_present(&ge))
1823 continue;
1824
1825 trace_spt_guest_change(vgpu->id, __func__, NULL,
1826 ge.type, ge.val64, index);
1827
1828 spt = ppgtt_populate_spt_by_guest_entry(vgpu, &ge);
1829 if (IS_ERR(spt)) {
1830 gvt_vgpu_err("fail to populate guest root pointer\n");
1831 ret = PTR_ERR(spt);
1832 goto fail;
1833 }
1834 ppgtt_generate_shadow_entry(&se, spt, &ge);
1835 ppgtt_set_shadow_root_entry(mm, &se, index);
1836
1837 trace_spt_guest_change(vgpu->id, "populate root pointer",
1838 NULL, se.type, se.val64, index);
1839 }
1840
1841 return 0;
1842 fail:
1843 invalidate_ppgtt_mm(mm);
1844 return ret;
1845 }
1846
1847 static struct intel_vgpu_mm *vgpu_alloc_mm(struct intel_vgpu *vgpu)
1848 {
1849 struct intel_vgpu_mm *mm;
1850
1851 mm = kzalloc(sizeof(*mm), GFP_KERNEL);
1852 if (!mm)
1853 return NULL;
1854
1855 mm->vgpu = vgpu;
1856 kref_init(&mm->ref);
1857 atomic_set(&mm->pincount, 0);
1858
1859 return mm;
1860 }
1861
1862 static void vgpu_free_mm(struct intel_vgpu_mm *mm)
1863 {
1864 kfree(mm);
1865 }
1866
1867 /**
1868 * intel_vgpu_create_ppgtt_mm - create a ppgtt mm object for a vGPU
1869 * @vgpu: a vGPU
1870 * @root_entry_type: ppgtt root entry type
1871 * @pdps: guest pdps.
1872 *
1873 * This function is used to create a ppgtt mm object for a vGPU.
1874 *
1875 * Returns:
1876 * Zero on success, negative error code in pointer if failed.
1877 */
1878 struct intel_vgpu_mm *intel_vgpu_create_ppgtt_mm(struct intel_vgpu *vgpu,
1879 enum intel_gvt_gtt_type root_entry_type, u64 pdps[])
1880 {
1881 struct intel_gvt *gvt = vgpu->gvt;
1882 struct intel_vgpu_mm *mm;
1883 int ret;
1884
1885 mm = vgpu_alloc_mm(vgpu);
1886 if (!mm)
1887 return ERR_PTR(-ENOMEM);
1888
1889 mm->type = INTEL_GVT_MM_PPGTT;
1890
1891 GEM_BUG_ON(root_entry_type != GTT_TYPE_PPGTT_ROOT_L3_ENTRY &&
1892 root_entry_type != GTT_TYPE_PPGTT_ROOT_L4_ENTRY);
1893 mm->ppgtt_mm.root_entry_type = root_entry_type;
1894
1895 INIT_LIST_HEAD(&mm->ppgtt_mm.list);
1896 INIT_LIST_HEAD(&mm->ppgtt_mm.lru_list);
1897
1898 if (root_entry_type == GTT_TYPE_PPGTT_ROOT_L4_ENTRY)
1899 mm->ppgtt_mm.guest_pdps[0] = pdps[0];
1900 else
1901 memcpy(mm->ppgtt_mm.guest_pdps, pdps,
1902 sizeof(mm->ppgtt_mm.guest_pdps));
1903
1904 ret = shadow_ppgtt_mm(mm);
1905 if (ret) {
1906 gvt_vgpu_err("failed to shadow ppgtt mm\n");
1907 vgpu_free_mm(mm);
1908 return ERR_PTR(ret);
1909 }
1910
1911 list_add_tail(&mm->ppgtt_mm.list, &vgpu->gtt.ppgtt_mm_list_head);
1912
1913 mutex_lock(&gvt->gtt.ppgtt_mm_lock);
1914 list_add_tail(&mm->ppgtt_mm.lru_list, &gvt->gtt.ppgtt_mm_lru_list_head);
1915 mutex_unlock(&gvt->gtt.ppgtt_mm_lock);
1916
1917 return mm;
1918 }
1919
1920 static struct intel_vgpu_mm *intel_vgpu_create_ggtt_mm(struct intel_vgpu *vgpu)
1921 {
1922 struct intel_vgpu_mm *mm;
1923 unsigned long nr_entries;
1924
1925 mm = vgpu_alloc_mm(vgpu);
1926 if (!mm)
1927 return ERR_PTR(-ENOMEM);
1928
1929 mm->type = INTEL_GVT_MM_GGTT;
1930
1931 nr_entries = gvt_ggtt_gm_sz(vgpu->gvt) >> I915_GTT_PAGE_SHIFT;
1932 mm->ggtt_mm.virtual_ggtt =
1933 vzalloc(array_size(nr_entries,
1934 vgpu->gvt->device_info.gtt_entry_size));
1935 if (!mm->ggtt_mm.virtual_ggtt) {
1936 vgpu_free_mm(mm);
1937 return ERR_PTR(-ENOMEM);
1938 }
1939
1940 return mm;
1941 }
1942
1943 /**
1944 * _intel_vgpu_mm_release - destroy a mm object
1945 * @mm_ref: a kref object
1946 *
1947 * This function is used to destroy a mm object for vGPU
1948 *
1949 */
1950 void _intel_vgpu_mm_release(struct kref *mm_ref)
1951 {
1952 struct intel_vgpu_mm *mm = container_of(mm_ref, typeof(*mm), ref);
1953
1954 if (GEM_WARN_ON(atomic_read(&mm->pincount)))
1955 gvt_err("vgpu mm pin count bug detected\n");
1956
1957 if (mm->type == INTEL_GVT_MM_PPGTT) {
1958 list_del(&mm->ppgtt_mm.list);
1959 list_del(&mm->ppgtt_mm.lru_list);
1960 invalidate_ppgtt_mm(mm);
1961 } else {
1962 vfree(mm->ggtt_mm.virtual_ggtt);
1963 }
1964
1965 vgpu_free_mm(mm);
1966 }
1967
1968 /**
1969 * intel_vgpu_unpin_mm - decrease the pin count of a vGPU mm object
1970 * @mm: a vGPU mm object
1971 *
1972 * This function is called when user doesn't want to use a vGPU mm object
1973 */
1974 void intel_vgpu_unpin_mm(struct intel_vgpu_mm *mm)
1975 {
1976 atomic_dec_if_positive(&mm->pincount);
1977 }
1978
1979 /**
1980 * intel_vgpu_pin_mm - increase the pin count of a vGPU mm object
1981 * @mm: target vgpu mm
1982 *
1983 * This function is called when user wants to use a vGPU mm object. If this
1984 * mm object hasn't been shadowed yet, the shadow will be populated at this
1985 * time.
1986 *
1987 * Returns:
1988 * Zero on success, negative error code if failed.
1989 */
1990 int intel_vgpu_pin_mm(struct intel_vgpu_mm *mm)
1991 {
1992 int ret;
1993
1994 atomic_inc(&mm->pincount);
1995
1996 if (mm->type == INTEL_GVT_MM_PPGTT) {
1997 ret = shadow_ppgtt_mm(mm);
1998 if (ret)
1999 return ret;
2000
2001 mutex_lock(&mm->vgpu->gvt->gtt.ppgtt_mm_lock);
2002 list_move_tail(&mm->ppgtt_mm.lru_list,
2003 &mm->vgpu->gvt->gtt.ppgtt_mm_lru_list_head);
2004 mutex_unlock(&mm->vgpu->gvt->gtt.ppgtt_mm_lock);
2005 }
2006
2007 return 0;
2008 }
2009
2010 static int reclaim_one_ppgtt_mm(struct intel_gvt *gvt)
2011 {
2012 struct intel_vgpu_mm *mm;
2013 struct list_head *pos, *n;
2014
2015 mutex_lock(&gvt->gtt.ppgtt_mm_lock);
2016
2017 list_for_each_safe(pos, n, &gvt->gtt.ppgtt_mm_lru_list_head) {
2018 mm = container_of(pos, struct intel_vgpu_mm, ppgtt_mm.lru_list);
2019
2020 if (atomic_read(&mm->pincount))
2021 continue;
2022
2023 list_del_init(&mm->ppgtt_mm.lru_list);
2024 mutex_unlock(&gvt->gtt.ppgtt_mm_lock);
2025 invalidate_ppgtt_mm(mm);
2026 return 1;
2027 }
2028 mutex_unlock(&gvt->gtt.ppgtt_mm_lock);
2029 return 0;
2030 }
2031
2032 /*
2033 * GMA translation APIs.
2034 */
2035 static inline int ppgtt_get_next_level_entry(struct intel_vgpu_mm *mm,
2036 struct intel_gvt_gtt_entry *e, unsigned long index, bool guest)
2037 {
2038 struct intel_vgpu *vgpu = mm->vgpu;
2039 struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
2040 struct intel_vgpu_ppgtt_spt *s;
2041
2042 s = intel_vgpu_find_spt_by_mfn(vgpu, ops->get_pfn(e));
2043 if (!s)
2044 return -ENXIO;
2045
2046 if (!guest)
2047 ppgtt_get_shadow_entry(s, e, index);
2048 else
2049 ppgtt_get_guest_entry(s, e, index);
2050 return 0;
2051 }
2052
2053 /**
2054 * intel_vgpu_gma_to_gpa - translate a gma to GPA
2055 * @mm: mm object. could be a PPGTT or GGTT mm object
2056 * @gma: graphics memory address in this mm object
2057 *
2058 * This function is used to translate a graphics memory address in specific
2059 * graphics memory space to guest physical address.
2060 *
2061 * Returns:
2062 * Guest physical address on success, INTEL_GVT_INVALID_ADDR if failed.
2063 */
2064 unsigned long intel_vgpu_gma_to_gpa(struct intel_vgpu_mm *mm, unsigned long gma)
2065 {
2066 struct intel_vgpu *vgpu = mm->vgpu;
2067 struct intel_gvt *gvt = vgpu->gvt;
2068 struct intel_gvt_gtt_pte_ops *pte_ops = gvt->gtt.pte_ops;
2069 struct intel_gvt_gtt_gma_ops *gma_ops = gvt->gtt.gma_ops;
2070 unsigned long gpa = INTEL_GVT_INVALID_ADDR;
2071 unsigned long gma_index[4];
2072 struct intel_gvt_gtt_entry e;
2073 int i, levels = 0;
2074 int ret;
2075
2076 GEM_BUG_ON(mm->type != INTEL_GVT_MM_GGTT &&
2077 mm->type != INTEL_GVT_MM_PPGTT);
2078
2079 if (mm->type == INTEL_GVT_MM_GGTT) {
2080 if (!vgpu_gmadr_is_valid(vgpu, gma))
2081 goto err;
2082
2083 ggtt_get_guest_entry(mm, &e,
2084 gma_ops->gma_to_ggtt_pte_index(gma));
2085
2086 gpa = (pte_ops->get_pfn(&e) << I915_GTT_PAGE_SHIFT)
2087 + (gma & ~I915_GTT_PAGE_MASK);
2088
2089 trace_gma_translate(vgpu->id, "ggtt", 0, 0, gma, gpa);
2090 } else {
2091 switch (mm->ppgtt_mm.root_entry_type) {
2092 case GTT_TYPE_PPGTT_ROOT_L4_ENTRY:
2093 ppgtt_get_shadow_root_entry(mm, &e, 0);
2094
2095 gma_index[0] = gma_ops->gma_to_pml4_index(gma);
2096 gma_index[1] = gma_ops->gma_to_l4_pdp_index(gma);
2097 gma_index[2] = gma_ops->gma_to_pde_index(gma);
2098 gma_index[3] = gma_ops->gma_to_pte_index(gma);
2099 levels = 4;
2100 break;
2101 case GTT_TYPE_PPGTT_ROOT_L3_ENTRY:
2102 ppgtt_get_shadow_root_entry(mm, &e,
2103 gma_ops->gma_to_l3_pdp_index(gma));
2104
2105 gma_index[0] = gma_ops->gma_to_pde_index(gma);
2106 gma_index[1] = gma_ops->gma_to_pte_index(gma);
2107 levels = 2;
2108 break;
2109 default:
2110 GEM_BUG_ON(1);
2111 }
2112
2113 /* walk the shadow page table and get gpa from guest entry */
2114 for (i = 0; i < levels; i++) {
2115 ret = ppgtt_get_next_level_entry(mm, &e, gma_index[i],
2116 (i == levels - 1));
2117 if (ret)
2118 goto err;
2119
2120 if (!pte_ops->test_present(&e)) {
2121 gvt_dbg_core("GMA 0x%lx is not present\n", gma);
2122 goto err;
2123 }
2124 }
2125
2126 gpa = (pte_ops->get_pfn(&e) << I915_GTT_PAGE_SHIFT) +
2127 (gma & ~I915_GTT_PAGE_MASK);
2128 trace_gma_translate(vgpu->id, "ppgtt", 0,
2129 mm->ppgtt_mm.root_entry_type, gma, gpa);
2130 }
2131
2132 return gpa;
2133 err:
2134 gvt_vgpu_err("invalid mm type: %d gma %lx\n", mm->type, gma);
2135 return INTEL_GVT_INVALID_ADDR;
2136 }
2137
2138 static int emulate_ggtt_mmio_read(struct intel_vgpu *vgpu,
2139 unsigned int off, void *p_data, unsigned int bytes)
2140 {
2141 struct intel_vgpu_mm *ggtt_mm = vgpu->gtt.ggtt_mm;
2142 const struct intel_gvt_device_info *info = &vgpu->gvt->device_info;
2143 unsigned long index = off >> info->gtt_entry_size_shift;
2144 unsigned long gma;
2145 struct intel_gvt_gtt_entry e;
2146
2147 if (bytes != 4 && bytes != 8)
2148 return -EINVAL;
2149
2150 gma = index << I915_GTT_PAGE_SHIFT;
2151 if (!intel_gvt_ggtt_validate_range(vgpu,
2152 gma, 1 << I915_GTT_PAGE_SHIFT)) {
2153 gvt_dbg_mm("read invalid ggtt at 0x%lx\n", gma);
2154 memset(p_data, 0, bytes);
2155 return 0;
2156 }
2157
2158 ggtt_get_guest_entry(ggtt_mm, &e, index);
2159 memcpy(p_data, (void *)&e.val64 + (off & (info->gtt_entry_size - 1)),
2160 bytes);
2161 return 0;
2162 }
2163
2164 /**
2165 * intel_vgpu_emulate_gtt_mmio_read - emulate GTT MMIO register read
2166 * @vgpu: a vGPU
2167 * @off: register offset
2168 * @p_data: data will be returned to guest
2169 * @bytes: data length
2170 *
2171 * This function is used to emulate the GTT MMIO register read
2172 *
2173 * Returns:
2174 * Zero on success, error code if failed.
2175 */
2176 int intel_vgpu_emulate_ggtt_mmio_read(struct intel_vgpu *vgpu, unsigned int off,
2177 void *p_data, unsigned int bytes)
2178 {
2179 const struct intel_gvt_device_info *info = &vgpu->gvt->device_info;
2180 int ret;
2181
2182 if (bytes != 4 && bytes != 8)
2183 return -EINVAL;
2184
2185 off -= info->gtt_start_offset;
2186 ret = emulate_ggtt_mmio_read(vgpu, off, p_data, bytes);
2187 return ret;
2188 }
2189
2190 static void ggtt_invalidate_pte(struct intel_vgpu *vgpu,
2191 struct intel_gvt_gtt_entry *entry)
2192 {
2193 struct intel_gvt_gtt_pte_ops *pte_ops = vgpu->gvt->gtt.pte_ops;
2194 unsigned long pfn;
2195
2196 pfn = pte_ops->get_pfn(entry);
2197 if (pfn != vgpu->gvt->gtt.scratch_mfn)
2198 intel_gvt_hypervisor_dma_unmap_guest_page(vgpu,
2199 pfn << PAGE_SHIFT);
2200 }
2201
2202 static int emulate_ggtt_mmio_write(struct intel_vgpu *vgpu, unsigned int off,
2203 void *p_data, unsigned int bytes)
2204 {
2205 struct intel_gvt *gvt = vgpu->gvt;
2206 const struct intel_gvt_device_info *info = &gvt->device_info;
2207 struct intel_vgpu_mm *ggtt_mm = vgpu->gtt.ggtt_mm;
2208 struct intel_gvt_gtt_pte_ops *ops = gvt->gtt.pte_ops;
2209 unsigned long g_gtt_index = off >> info->gtt_entry_size_shift;
2210 unsigned long gma, gfn;
2211 struct intel_gvt_gtt_entry e = {.val64 = 0, .type = GTT_TYPE_GGTT_PTE};
2212 struct intel_gvt_gtt_entry m = {.val64 = 0, .type = GTT_TYPE_GGTT_PTE};
2213 dma_addr_t dma_addr;
2214 int ret;
2215 struct intel_gvt_partial_pte *partial_pte, *pos, *n;
2216 bool partial_update = false;
2217
2218 if (bytes != 4 && bytes != 8)
2219 return -EINVAL;
2220
2221 gma = g_gtt_index << I915_GTT_PAGE_SHIFT;
2222
2223 /* the VM may configure the whole GM space when ballooning is used */
2224 if (!vgpu_gmadr_is_valid(vgpu, gma))
2225 return 0;
2226
2227 e.type = GTT_TYPE_GGTT_PTE;
2228 memcpy((void *)&e.val64 + (off & (info->gtt_entry_size - 1)), p_data,
2229 bytes);
2230
2231 /* If ggtt entry size is 8 bytes, and it's split into two 4 bytes
2232 * write, save the first 4 bytes in a list and update virtual
2233 * PTE. Only update shadow PTE when the second 4 bytes comes.
2234 */
2235 if (bytes < info->gtt_entry_size) {
2236 bool found = false;
2237
2238 list_for_each_entry_safe(pos, n,
2239 &ggtt_mm->ggtt_mm.partial_pte_list, list) {
2240 if (g_gtt_index == pos->offset >>
2241 info->gtt_entry_size_shift) {
2242 if (off != pos->offset) {
2243 /* the second partial part*/
2244 int last_off = pos->offset &
2245 (info->gtt_entry_size - 1);
2246
2247 memcpy((void *)&e.val64 + last_off,
2248 (void *)&pos->data + last_off,
2249 bytes);
2250
2251 list_del(&pos->list);
2252 kfree(pos);
2253 found = true;
2254 break;
2255 }
2256
2257 /* update of the first partial part */
2258 pos->data = e.val64;
2259 ggtt_set_guest_entry(ggtt_mm, &e, g_gtt_index);
2260 return 0;
2261 }
2262 }
2263
2264 if (!found) {
2265 /* the first partial part */
2266 partial_pte = kzalloc(sizeof(*partial_pte), GFP_KERNEL);
2267 if (!partial_pte)
2268 return -ENOMEM;
2269 partial_pte->offset = off;
2270 partial_pte->data = e.val64;
2271 list_add_tail(&partial_pte->list,
2272 &ggtt_mm->ggtt_mm.partial_pte_list);
2273 partial_update = true;
2274 }
2275 }
2276
2277 if (!partial_update && (ops->test_present(&e))) {
2278 gfn = ops->get_pfn(&e);
2279 m.val64 = e.val64;
2280 m.type = e.type;
2281
2282 /* one PTE update may be issued in multiple writes and the
2283 * first write may not construct a valid gfn
2284 */
2285 if (!intel_gvt_hypervisor_is_valid_gfn(vgpu, gfn)) {
2286 ops->set_pfn(&m, gvt->gtt.scratch_mfn);
2287 goto out;
2288 }
2289
2290 ret = intel_gvt_hypervisor_dma_map_guest_page(vgpu, gfn,
2291 PAGE_SIZE, &dma_addr);
2292 if (ret) {
2293 gvt_vgpu_err("fail to populate guest ggtt entry\n");
2294 /* guest driver may read/write the entry when partial
2295 * update the entry in this situation p2m will fail
2296 * settting the shadow entry to point to a scratch page
2297 */
2298 ops->set_pfn(&m, gvt->gtt.scratch_mfn);
2299 } else
2300 ops->set_pfn(&m, dma_addr >> PAGE_SHIFT);
2301 } else {
2302 ops->set_pfn(&m, gvt->gtt.scratch_mfn);
2303 ops->clear_present(&m);
2304 }
2305
2306 out:
2307 ggtt_set_guest_entry(ggtt_mm, &e, g_gtt_index);
2308
2309 ggtt_get_host_entry(ggtt_mm, &e, g_gtt_index);
2310 ggtt_invalidate_pte(vgpu, &e);
2311
2312 ggtt_set_host_entry(ggtt_mm, &m, g_gtt_index);
2313 ggtt_invalidate(gvt->dev_priv);
2314 return 0;
2315 }
2316
2317 /*
2318 * intel_vgpu_emulate_ggtt_mmio_write - emulate GTT MMIO register write
2319 * @vgpu: a vGPU
2320 * @off: register offset
2321 * @p_data: data from guest write
2322 * @bytes: data length
2323 *
2324 * This function is used to emulate the GTT MMIO register write
2325 *
2326 * Returns:
2327 * Zero on success, error code if failed.
2328 */
2329 int intel_vgpu_emulate_ggtt_mmio_write(struct intel_vgpu *vgpu,
2330 unsigned int off, void *p_data, unsigned int bytes)
2331 {
2332 const struct intel_gvt_device_info *info = &vgpu->gvt->device_info;
2333 int ret;
2334
2335 if (bytes != 4 && bytes != 8)
2336 return -EINVAL;
2337
2338 off -= info->gtt_start_offset;
2339 ret = emulate_ggtt_mmio_write(vgpu, off, p_data, bytes);
2340 return ret;
2341 }
2342
2343 static int alloc_scratch_pages(struct intel_vgpu *vgpu,
2344 enum intel_gvt_gtt_type type)
2345 {
2346 struct intel_vgpu_gtt *gtt = &vgpu->gtt;
2347 struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
2348 int page_entry_num = I915_GTT_PAGE_SIZE >>
2349 vgpu->gvt->device_info.gtt_entry_size_shift;
2350 void *scratch_pt;
2351 int i;
2352 struct device *dev = &vgpu->gvt->dev_priv->drm.pdev->dev;
2353 dma_addr_t daddr;
2354
2355 if (WARN_ON(type < GTT_TYPE_PPGTT_PTE_PT || type >= GTT_TYPE_MAX))
2356 return -EINVAL;
2357
2358 scratch_pt = (void *)get_zeroed_page(GFP_KERNEL);
2359 if (!scratch_pt) {
2360 gvt_vgpu_err("fail to allocate scratch page\n");
2361 return -ENOMEM;
2362 }
2363
2364 daddr = dma_map_page(dev, virt_to_page(scratch_pt), 0,
2365 4096, PCI_DMA_BIDIRECTIONAL);
2366 if (dma_mapping_error(dev, daddr)) {
2367 gvt_vgpu_err("fail to dmamap scratch_pt\n");
2368 __free_page(virt_to_page(scratch_pt));
2369 return -ENOMEM;
2370 }
2371 gtt->scratch_pt[type].page_mfn =
2372 (unsigned long)(daddr >> I915_GTT_PAGE_SHIFT);
2373 gtt->scratch_pt[type].page = virt_to_page(scratch_pt);
2374 gvt_dbg_mm("vgpu%d create scratch_pt: type %d mfn=0x%lx\n",
2375 vgpu->id, type, gtt->scratch_pt[type].page_mfn);
2376
2377 /* Build the tree by full filled the scratch pt with the entries which
2378 * point to the next level scratch pt or scratch page. The
2379 * scratch_pt[type] indicate the scratch pt/scratch page used by the
2380 * 'type' pt.
2381 * e.g. scratch_pt[GTT_TYPE_PPGTT_PDE_PT] is used by
2382 * GTT_TYPE_PPGTT_PDE_PT level pt, that means this scratch_pt it self
2383 * is GTT_TYPE_PPGTT_PTE_PT, and full filled by scratch page mfn.
2384 */
2385 if (type > GTT_TYPE_PPGTT_PTE_PT) {
2386 struct intel_gvt_gtt_entry se;
2387
2388 memset(&se, 0, sizeof(struct intel_gvt_gtt_entry));
2389 se.type = get_entry_type(type - 1);
2390 ops->set_pfn(&se, gtt->scratch_pt[type - 1].page_mfn);
2391
2392 /* The entry parameters like present/writeable/cache type
2393 * set to the same as i915's scratch page tree.
2394 */
2395 se.val64 |= _PAGE_PRESENT | _PAGE_RW;
2396 if (type == GTT_TYPE_PPGTT_PDE_PT)
2397 se.val64 |= PPAT_CACHED;
2398
2399 for (i = 0; i < page_entry_num; i++)
2400 ops->set_entry(scratch_pt, &se, i, false, 0, vgpu);
2401 }
2402
2403 return 0;
2404 }
2405
2406 static int release_scratch_page_tree(struct intel_vgpu *vgpu)
2407 {
2408 int i;
2409 struct device *dev = &vgpu->gvt->dev_priv->drm.pdev->dev;
2410 dma_addr_t daddr;
2411
2412 for (i = GTT_TYPE_PPGTT_PTE_PT; i < GTT_TYPE_MAX; i++) {
2413 if (vgpu->gtt.scratch_pt[i].page != NULL) {
2414 daddr = (dma_addr_t)(vgpu->gtt.scratch_pt[i].page_mfn <<
2415 I915_GTT_PAGE_SHIFT);
2416 dma_unmap_page(dev, daddr, 4096, PCI_DMA_BIDIRECTIONAL);
2417 __free_page(vgpu->gtt.scratch_pt[i].page);
2418 vgpu->gtt.scratch_pt[i].page = NULL;
2419 vgpu->gtt.scratch_pt[i].page_mfn = 0;
2420 }
2421 }
2422
2423 return 0;
2424 }
2425
2426 static int create_scratch_page_tree(struct intel_vgpu *vgpu)
2427 {
2428 int i, ret;
2429
2430 for (i = GTT_TYPE_PPGTT_PTE_PT; i < GTT_TYPE_MAX; i++) {
2431 ret = alloc_scratch_pages(vgpu, i);
2432 if (ret)
2433 goto err;
2434 }
2435
2436 return 0;
2437
2438 err:
2439 release_scratch_page_tree(vgpu);
2440 return ret;
2441 }
2442
2443 /**
2444 * intel_vgpu_init_gtt - initialize per-vGPU graphics memory virulization
2445 * @vgpu: a vGPU
2446 *
2447 * This function is used to initialize per-vGPU graphics memory virtualization
2448 * components.
2449 *
2450 * Returns:
2451 * Zero on success, error code if failed.
2452 */
2453 int intel_vgpu_init_gtt(struct intel_vgpu *vgpu)
2454 {
2455 struct intel_vgpu_gtt *gtt = &vgpu->gtt;
2456
2457 INIT_RADIX_TREE(&gtt->spt_tree, GFP_KERNEL);
2458
2459 INIT_LIST_HEAD(&gtt->ppgtt_mm_list_head);
2460 INIT_LIST_HEAD(&gtt->oos_page_list_head);
2461 INIT_LIST_HEAD(&gtt->post_shadow_list_head);
2462
2463 gtt->ggtt_mm = intel_vgpu_create_ggtt_mm(vgpu);
2464 if (IS_ERR(gtt->ggtt_mm)) {
2465 gvt_vgpu_err("fail to create mm for ggtt.\n");
2466 return PTR_ERR(gtt->ggtt_mm);
2467 }
2468
2469 intel_vgpu_reset_ggtt(vgpu, false);
2470
2471 INIT_LIST_HEAD(&gtt->ggtt_mm->ggtt_mm.partial_pte_list);
2472
2473 return create_scratch_page_tree(vgpu);
2474 }
2475
2476 static void intel_vgpu_destroy_all_ppgtt_mm(struct intel_vgpu *vgpu)
2477 {
2478 struct list_head *pos, *n;
2479 struct intel_vgpu_mm *mm;
2480
2481 list_for_each_safe(pos, n, &vgpu->gtt.ppgtt_mm_list_head) {
2482 mm = container_of(pos, struct intel_vgpu_mm, ppgtt_mm.list);
2483 intel_vgpu_destroy_mm(mm);
2484 }
2485
2486 if (GEM_WARN_ON(!list_empty(&vgpu->gtt.ppgtt_mm_list_head)))
2487 gvt_err("vgpu ppgtt mm is not fully destroyed\n");
2488
2489 if (GEM_WARN_ON(!radix_tree_empty(&vgpu->gtt.spt_tree))) {
2490 gvt_err("Why we still has spt not freed?\n");
2491 ppgtt_free_all_spt(vgpu);
2492 }
2493 }
2494
2495 static void intel_vgpu_destroy_ggtt_mm(struct intel_vgpu *vgpu)
2496 {
2497 struct intel_gvt_partial_pte *pos, *next;
2498
2499 list_for_each_entry_safe(pos, next,
2500 &vgpu->gtt.ggtt_mm->ggtt_mm.partial_pte_list,
2501 list) {
2502 gvt_dbg_mm("partial PTE update on hold 0x%lx : 0x%llx\n",
2503 pos->offset, pos->data);
2504 kfree(pos);
2505 }
2506 intel_vgpu_destroy_mm(vgpu->gtt.ggtt_mm);
2507 vgpu->gtt.ggtt_mm = NULL;
2508 }
2509
2510 /**
2511 * intel_vgpu_clean_gtt - clean up per-vGPU graphics memory virulization
2512 * @vgpu: a vGPU
2513 *
2514 * This function is used to clean up per-vGPU graphics memory virtualization
2515 * components.
2516 *
2517 * Returns:
2518 * Zero on success, error code if failed.
2519 */
2520 void intel_vgpu_clean_gtt(struct intel_vgpu *vgpu)
2521 {
2522 intel_vgpu_destroy_all_ppgtt_mm(vgpu);
2523 intel_vgpu_destroy_ggtt_mm(vgpu);
2524 release_scratch_page_tree(vgpu);
2525 }
2526
2527 static void clean_spt_oos(struct intel_gvt *gvt)
2528 {
2529 struct intel_gvt_gtt *gtt = &gvt->gtt;
2530 struct list_head *pos, *n;
2531 struct intel_vgpu_oos_page *oos_page;
2532
2533 WARN(!list_empty(&gtt->oos_page_use_list_head),
2534 "someone is still using oos page\n");
2535
2536 list_for_each_safe(pos, n, &gtt->oos_page_free_list_head) {
2537 oos_page = container_of(pos, struct intel_vgpu_oos_page, list);
2538 list_del(&oos_page->list);
2539 free_page((unsigned long)oos_page->mem);
2540 kfree(oos_page);
2541 }
2542 }
2543
2544 static int setup_spt_oos(struct intel_gvt *gvt)
2545 {
2546 struct intel_gvt_gtt *gtt = &gvt->gtt;
2547 struct intel_vgpu_oos_page *oos_page;
2548 int i;
2549 int ret;
2550
2551 INIT_LIST_HEAD(&gtt->oos_page_free_list_head);
2552 INIT_LIST_HEAD(&gtt->oos_page_use_list_head);
2553
2554 for (i = 0; i < preallocated_oos_pages; i++) {
2555 oos_page = kzalloc(sizeof(*oos_page), GFP_KERNEL);
2556 if (!oos_page) {
2557 ret = -ENOMEM;
2558 goto fail;
2559 }
2560 oos_page->mem = (void *)__get_free_pages(GFP_KERNEL, 0);
2561 if (!oos_page->mem) {
2562 ret = -ENOMEM;
2563 kfree(oos_page);
2564 goto fail;
2565 }
2566
2567 INIT_LIST_HEAD(&oos_page->list);
2568 INIT_LIST_HEAD(&oos_page->vm_list);
2569 oos_page->id = i;
2570 list_add_tail(&oos_page->list, &gtt->oos_page_free_list_head);
2571 }
2572
2573 gvt_dbg_mm("%d oos pages preallocated\n", i);
2574
2575 return 0;
2576 fail:
2577 clean_spt_oos(gvt);
2578 return ret;
2579 }
2580
2581 /**
2582 * intel_vgpu_find_ppgtt_mm - find a PPGTT mm object
2583 * @vgpu: a vGPU
2584 * @pdps: pdp root array
2585 *
2586 * This function is used to find a PPGTT mm object from mm object pool
2587 *
2588 * Returns:
2589 * pointer to mm object on success, NULL if failed.
2590 */
2591 struct intel_vgpu_mm *intel_vgpu_find_ppgtt_mm(struct intel_vgpu *vgpu,
2592 u64 pdps[])
2593 {
2594 struct intel_vgpu_mm *mm;
2595 struct list_head *pos;
2596
2597 list_for_each(pos, &vgpu->gtt.ppgtt_mm_list_head) {
2598 mm = container_of(pos, struct intel_vgpu_mm, ppgtt_mm.list);
2599
2600 switch (mm->ppgtt_mm.root_entry_type) {
2601 case GTT_TYPE_PPGTT_ROOT_L4_ENTRY:
2602 if (pdps[0] == mm->ppgtt_mm.guest_pdps[0])
2603 return mm;
2604 break;
2605 case GTT_TYPE_PPGTT_ROOT_L3_ENTRY:
2606 if (!memcmp(pdps, mm->ppgtt_mm.guest_pdps,
2607 sizeof(mm->ppgtt_mm.guest_pdps)))
2608 return mm;
2609 break;
2610 default:
2611 GEM_BUG_ON(1);
2612 }
2613 }
2614 return NULL;
2615 }
2616
2617 /**
2618 * intel_vgpu_get_ppgtt_mm - get or create a PPGTT mm object.
2619 * @vgpu: a vGPU
2620 * @root_entry_type: ppgtt root entry type
2621 * @pdps: guest pdps
2622 *
2623 * This function is used to find or create a PPGTT mm object from a guest.
2624 *
2625 * Returns:
2626 * Zero on success, negative error code if failed.
2627 */
2628 struct intel_vgpu_mm *intel_vgpu_get_ppgtt_mm(struct intel_vgpu *vgpu,
2629 enum intel_gvt_gtt_type root_entry_type, u64 pdps[])
2630 {
2631 struct intel_vgpu_mm *mm;
2632
2633 mm = intel_vgpu_find_ppgtt_mm(vgpu, pdps);
2634 if (mm) {
2635 intel_vgpu_mm_get(mm);
2636 } else {
2637 mm = intel_vgpu_create_ppgtt_mm(vgpu, root_entry_type, pdps);
2638 if (IS_ERR(mm))
2639 gvt_vgpu_err("fail to create mm\n");
2640 }
2641 return mm;
2642 }
2643
2644 /**
2645 * intel_vgpu_put_ppgtt_mm - find and put a PPGTT mm object.
2646 * @vgpu: a vGPU
2647 * @pdps: guest pdps
2648 *
2649 * This function is used to find a PPGTT mm object from a guest and destroy it.
2650 *
2651 * Returns:
2652 * Zero on success, negative error code if failed.
2653 */
2654 int intel_vgpu_put_ppgtt_mm(struct intel_vgpu *vgpu, u64 pdps[])
2655 {
2656 struct intel_vgpu_mm *mm;
2657
2658 mm = intel_vgpu_find_ppgtt_mm(vgpu, pdps);
2659 if (!mm) {
2660 gvt_vgpu_err("fail to find ppgtt instance.\n");
2661 return -EINVAL;
2662 }
2663 intel_vgpu_mm_put(mm);
2664 return 0;
2665 }
2666
2667 /**
2668 * intel_gvt_init_gtt - initialize mm components of a GVT device
2669 * @gvt: GVT device
2670 *
2671 * This function is called at the initialization stage, to initialize
2672 * the mm components of a GVT device.
2673 *
2674 * Returns:
2675 * zero on success, negative error code if failed.
2676 */
2677 int intel_gvt_init_gtt(struct intel_gvt *gvt)
2678 {
2679 int ret;
2680 void *page;
2681 struct device *dev = &gvt->dev_priv->drm.pdev->dev;
2682 dma_addr_t daddr;
2683
2684 gvt_dbg_core("init gtt\n");
2685
2686 gvt->gtt.pte_ops = &gen8_gtt_pte_ops;
2687 gvt->gtt.gma_ops = &gen8_gtt_gma_ops;
2688
2689 page = (void *)get_zeroed_page(GFP_KERNEL);
2690 if (!page) {
2691 gvt_err("fail to allocate scratch ggtt page\n");
2692 return -ENOMEM;
2693 }
2694
2695 daddr = dma_map_page(dev, virt_to_page(page), 0,
2696 4096, PCI_DMA_BIDIRECTIONAL);
2697 if (dma_mapping_error(dev, daddr)) {
2698 gvt_err("fail to dmamap scratch ggtt page\n");
2699 __free_page(virt_to_page(page));
2700 return -ENOMEM;
2701 }
2702
2703 gvt->gtt.scratch_page = virt_to_page(page);
2704 gvt->gtt.scratch_mfn = (unsigned long)(daddr >> I915_GTT_PAGE_SHIFT);
2705
2706 if (enable_out_of_sync) {
2707 ret = setup_spt_oos(gvt);
2708 if (ret) {
2709 gvt_err("fail to initialize SPT oos\n");
2710 dma_unmap_page(dev, daddr, 4096, PCI_DMA_BIDIRECTIONAL);
2711 __free_page(gvt->gtt.scratch_page);
2712 return ret;
2713 }
2714 }
2715 INIT_LIST_HEAD(&gvt->gtt.ppgtt_mm_lru_list_head);
2716 mutex_init(&gvt->gtt.ppgtt_mm_lock);
2717 return 0;
2718 }
2719
2720 /**
2721 * intel_gvt_clean_gtt - clean up mm components of a GVT device
2722 * @gvt: GVT device
2723 *
2724 * This function is called at the driver unloading stage, to clean up the
2725 * the mm components of a GVT device.
2726 *
2727 */
2728 void intel_gvt_clean_gtt(struct intel_gvt *gvt)
2729 {
2730 struct device *dev = &gvt->dev_priv->drm.pdev->dev;
2731 dma_addr_t daddr = (dma_addr_t)(gvt->gtt.scratch_mfn <<
2732 I915_GTT_PAGE_SHIFT);
2733
2734 dma_unmap_page(dev, daddr, 4096, PCI_DMA_BIDIRECTIONAL);
2735
2736 __free_page(gvt->gtt.scratch_page);
2737
2738 if (enable_out_of_sync)
2739 clean_spt_oos(gvt);
2740 }
2741
2742 /**
2743 * intel_vgpu_invalidate_ppgtt - invalidate PPGTT instances
2744 * @vgpu: a vGPU
2745 *
2746 * This function is called when invalidate all PPGTT instances of a vGPU.
2747 *
2748 */
2749 void intel_vgpu_invalidate_ppgtt(struct intel_vgpu *vgpu)
2750 {
2751 struct list_head *pos, *n;
2752 struct intel_vgpu_mm *mm;
2753
2754 list_for_each_safe(pos, n, &vgpu->gtt.ppgtt_mm_list_head) {
2755 mm = container_of(pos, struct intel_vgpu_mm, ppgtt_mm.list);
2756 if (mm->type == INTEL_GVT_MM_PPGTT) {
2757 mutex_lock(&vgpu->gvt->gtt.ppgtt_mm_lock);
2758 list_del_init(&mm->ppgtt_mm.lru_list);
2759 mutex_unlock(&vgpu->gvt->gtt.ppgtt_mm_lock);
2760 if (mm->ppgtt_mm.shadowed)
2761 invalidate_ppgtt_mm(mm);
2762 }
2763 }
2764 }
2765
2766 /**
2767 * intel_vgpu_reset_ggtt - reset the GGTT entry
2768 * @vgpu: a vGPU
2769 * @invalidate_old: invalidate old entries
2770 *
2771 * This function is called at the vGPU create stage
2772 * to reset all the GGTT entries.
2773 *
2774 */
2775 void intel_vgpu_reset_ggtt(struct intel_vgpu *vgpu, bool invalidate_old)
2776 {
2777 struct intel_gvt *gvt = vgpu->gvt;
2778 struct drm_i915_private *dev_priv = gvt->dev_priv;
2779 struct intel_gvt_gtt_pte_ops *pte_ops = vgpu->gvt->gtt.pte_ops;
2780 struct intel_gvt_gtt_entry entry = {.type = GTT_TYPE_GGTT_PTE};
2781 struct intel_gvt_gtt_entry old_entry;
2782 u32 index;
2783 u32 num_entries;
2784
2785 pte_ops->set_pfn(&entry, gvt->gtt.scratch_mfn);
2786 pte_ops->set_present(&entry);
2787
2788 index = vgpu_aperture_gmadr_base(vgpu) >> PAGE_SHIFT;
2789 num_entries = vgpu_aperture_sz(vgpu) >> PAGE_SHIFT;
2790 while (num_entries--) {
2791 if (invalidate_old) {
2792 ggtt_get_host_entry(vgpu->gtt.ggtt_mm, &old_entry, index);
2793 ggtt_invalidate_pte(vgpu, &old_entry);
2794 }
2795 ggtt_set_host_entry(vgpu->gtt.ggtt_mm, &entry, index++);
2796 }
2797
2798 index = vgpu_hidden_gmadr_base(vgpu) >> PAGE_SHIFT;
2799 num_entries = vgpu_hidden_sz(vgpu) >> PAGE_SHIFT;
2800 while (num_entries--) {
2801 if (invalidate_old) {
2802 ggtt_get_host_entry(vgpu->gtt.ggtt_mm, &old_entry, index);
2803 ggtt_invalidate_pte(vgpu, &old_entry);
2804 }
2805 ggtt_set_host_entry(vgpu->gtt.ggtt_mm, &entry, index++);
2806 }
2807
2808 ggtt_invalidate(dev_priv);
2809 }
2810
2811 /**
2812 * intel_vgpu_reset_gtt - reset the all GTT related status
2813 * @vgpu: a vGPU
2814 *
2815 * This function is called from vfio core to reset reset all
2816 * GTT related status, including GGTT, PPGTT, scratch page.
2817 *
2818 */
2819 void intel_vgpu_reset_gtt(struct intel_vgpu *vgpu)
2820 {
2821 /* Shadow pages are only created when there is no page
2822 * table tracking data, so remove page tracking data after
2823 * removing the shadow pages.
2824 */
2825 intel_vgpu_destroy_all_ppgtt_mm(vgpu);
2826 intel_vgpu_reset_ggtt(vgpu, true);
2827 }
Linux with added FPC-III support