1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2020 Intel Corporation
4  */
5 
6 #include <linux/slab.h>
7 
8 #include "gem/i915_gem_lmem.h"
9 
10 #include "i915_trace.h"
11 #include "intel_gtt.h"
12 #include "gen6_ppgtt.h"
13 #include "gen8_ppgtt.h"
14 
alloc_pt(struct i915_address_space * vm)15 struct i915_page_table *alloc_pt(struct i915_address_space *vm)
16 {
17 	struct i915_page_table *pt;
18 
19 	pt = kmalloc(sizeof(*pt), I915_GFP_ALLOW_FAIL);
20 	if (unlikely(!pt))
21 		return ERR_PTR(-ENOMEM);
22 
23 	pt->base = vm->alloc_pt_dma(vm, I915_GTT_PAGE_SIZE_4K);
24 	if (IS_ERR(pt->base)) {
25 		kfree(pt);
26 		return ERR_PTR(-ENOMEM);
27 	}
28 
29 	atomic_set(&pt->used, 0);
30 	return pt;
31 }
32 
__alloc_pd(int count)33 struct i915_page_directory *__alloc_pd(int count)
34 {
35 	struct i915_page_directory *pd;
36 
37 	pd = kzalloc(sizeof(*pd), I915_GFP_ALLOW_FAIL);
38 	if (unlikely(!pd))
39 		return NULL;
40 
41 	pd->entry = kcalloc(count, sizeof(*pd->entry), I915_GFP_ALLOW_FAIL);
42 	if (unlikely(!pd->entry)) {
43 		kfree(pd);
44 		return NULL;
45 	}
46 
47 	spin_lock_init(&pd->lock);
48 	return pd;
49 }
50 
alloc_pd(struct i915_address_space * vm)51 struct i915_page_directory *alloc_pd(struct i915_address_space *vm)
52 {
53 	struct i915_page_directory *pd;
54 
55 	pd = __alloc_pd(I915_PDES);
56 	if (unlikely(!pd))
57 		return ERR_PTR(-ENOMEM);
58 
59 	pd->pt.base = vm->alloc_pt_dma(vm, I915_GTT_PAGE_SIZE_4K);
60 	if (IS_ERR(pd->pt.base)) {
61 		kfree(pd->entry);
62 		kfree(pd);
63 		return ERR_PTR(-ENOMEM);
64 	}
65 
66 	return pd;
67 }
68 
free_px(struct i915_address_space * vm,struct i915_page_table * pt,int lvl)69 void free_px(struct i915_address_space *vm, struct i915_page_table *pt, int lvl)
70 {
71 	BUILD_BUG_ON(offsetof(struct i915_page_directory, pt));
72 
73 	if (lvl) {
74 		struct i915_page_directory *pd =
75 			container_of(pt, typeof(*pd), pt);
76 		kfree(pd->entry);
77 	}
78 
79 	if (pt->base)
80 		i915_gem_object_put(pt->base);
81 
82 	kfree(pt);
83 }
84 
85 static void
write_dma_entry(struct drm_i915_gem_object * const pdma,const unsigned short idx,const u64 encoded_entry)86 write_dma_entry(struct drm_i915_gem_object * const pdma,
87 		const unsigned short idx,
88 		const u64 encoded_entry)
89 {
90 	u64 * const vaddr = __px_vaddr(pdma);
91 
92 	vaddr[idx] = encoded_entry;
93 	clflush_cache_range(&vaddr[idx], sizeof(u64));
94 }
95 
96 void
__set_pd_entry(struct i915_page_directory * const pd,const unsigned short idx,struct i915_page_table * const to,u64 (* encode)(const dma_addr_t,const enum i915_cache_level))97 __set_pd_entry(struct i915_page_directory * const pd,
98 	       const unsigned short idx,
99 	       struct i915_page_table * const to,
100 	       u64 (*encode)(const dma_addr_t, const enum i915_cache_level))
101 {
102 	/* Each thread pre-pins the pd, and we may have a thread per pde. */
103 	GEM_BUG_ON(atomic_read(px_used(pd)) > NALLOC * I915_PDES);
104 
105 	atomic_inc(px_used(pd));
106 	pd->entry[idx] = to;
107 	write_dma_entry(px_base(pd), idx, encode(px_dma(to), I915_CACHE_LLC));
108 }
109 
110 void
clear_pd_entry(struct i915_page_directory * const pd,const unsigned short idx,const struct drm_i915_gem_object * const scratch)111 clear_pd_entry(struct i915_page_directory * const pd,
112 	       const unsigned short idx,
113 	       const struct drm_i915_gem_object * const scratch)
114 {
115 	GEM_BUG_ON(atomic_read(px_used(pd)) == 0);
116 
117 	write_dma_entry(px_base(pd), idx, scratch->encode);
118 	pd->entry[idx] = NULL;
119 	atomic_dec(px_used(pd));
120 }
121 
122 bool
release_pd_entry(struct i915_page_directory * const pd,const unsigned short idx,struct i915_page_table * const pt,const struct drm_i915_gem_object * const scratch)123 release_pd_entry(struct i915_page_directory * const pd,
124 		 const unsigned short idx,
125 		 struct i915_page_table * const pt,
126 		 const struct drm_i915_gem_object * const scratch)
127 {
128 	bool free = false;
129 
130 	if (atomic_add_unless(&pt->used, -1, 1))
131 		return false;
132 
133 	spin_lock(&pd->lock);
134 	if (atomic_dec_and_test(&pt->used)) {
135 		clear_pd_entry(pd, idx, scratch);
136 		free = true;
137 	}
138 	spin_unlock(&pd->lock);
139 
140 	return free;
141 }
142 
i915_ppgtt_init_hw(struct intel_gt * gt)143 int i915_ppgtt_init_hw(struct intel_gt *gt)
144 {
145 	struct drm_i915_private *i915 = gt->i915;
146 
147 	gtt_write_workarounds(gt);
148 
149 	if (GRAPHICS_VER(i915) == 6)
150 		gen6_ppgtt_enable(gt);
151 	else if (GRAPHICS_VER(i915) == 7)
152 		gen7_ppgtt_enable(gt);
153 
154 	return 0;
155 }
156 
157 static struct i915_ppgtt *
__ppgtt_create(struct intel_gt * gt)158 __ppgtt_create(struct intel_gt *gt)
159 {
160 	if (GRAPHICS_VER(gt->i915) < 8)
161 		return gen6_ppgtt_create(gt);
162 	else
163 		return gen8_ppgtt_create(gt);
164 }
165 
i915_ppgtt_create(struct intel_gt * gt)166 struct i915_ppgtt *i915_ppgtt_create(struct intel_gt *gt)
167 {
168 	struct i915_ppgtt *ppgtt;
169 
170 	ppgtt = __ppgtt_create(gt);
171 	if (IS_ERR(ppgtt))
172 		return ppgtt;
173 
174 	trace_i915_ppgtt_create(&ppgtt->vm);
175 
176 	return ppgtt;
177 }
178 
ppgtt_bind_vma(struct i915_address_space * vm,struct i915_vm_pt_stash * stash,struct i915_vma * vma,enum i915_cache_level cache_level,u32 flags)179 void ppgtt_bind_vma(struct i915_address_space *vm,
180 		    struct i915_vm_pt_stash *stash,
181 		    struct i915_vma *vma,
182 		    enum i915_cache_level cache_level,
183 		    u32 flags)
184 {
185 	u32 pte_flags;
186 
187 	if (!test_bit(I915_VMA_ALLOC_BIT, __i915_vma_flags(vma))) {
188 		vm->allocate_va_range(vm, stash, vma->node.start, vma->size);
189 		set_bit(I915_VMA_ALLOC_BIT, __i915_vma_flags(vma));
190 	}
191 
192 	/* Applicable to VLV, and gen8+ */
193 	pte_flags = 0;
194 	if (i915_gem_object_is_readonly(vma->obj))
195 		pte_flags |= PTE_READ_ONLY;
196 	if (i915_gem_object_is_lmem(vma->obj))
197 		pte_flags |= PTE_LM;
198 
199 	vm->insert_entries(vm, vma, cache_level, pte_flags);
200 	wmb();
201 }
202 
ppgtt_unbind_vma(struct i915_address_space * vm,struct i915_vma * vma)203 void ppgtt_unbind_vma(struct i915_address_space *vm, struct i915_vma *vma)
204 {
205 	if (test_and_clear_bit(I915_VMA_ALLOC_BIT, __i915_vma_flags(vma)))
206 		vm->clear_range(vm, vma->node.start, vma->size);
207 }
208 
pd_count(u64 size,int shift)209 static unsigned long pd_count(u64 size, int shift)
210 {
211 	/* Beware later misalignment */
212 	return (size + 2 * (BIT_ULL(shift) - 1)) >> shift;
213 }
214 
i915_vm_alloc_pt_stash(struct i915_address_space * vm,struct i915_vm_pt_stash * stash,u64 size)215 int i915_vm_alloc_pt_stash(struct i915_address_space *vm,
216 			   struct i915_vm_pt_stash *stash,
217 			   u64 size)
218 {
219 	unsigned long count;
220 	int shift, n;
221 
222 	shift = vm->pd_shift;
223 	if (!shift)
224 		return 0;
225 
226 	count = pd_count(size, shift);
227 	while (count--) {
228 		struct i915_page_table *pt;
229 
230 		pt = alloc_pt(vm);
231 		if (IS_ERR(pt)) {
232 			i915_vm_free_pt_stash(vm, stash);
233 			return PTR_ERR(pt);
234 		}
235 
236 		pt->stash = stash->pt[0];
237 		stash->pt[0] = pt;
238 	}
239 
240 	for (n = 1; n < vm->top; n++) {
241 		shift += ilog2(I915_PDES); /* Each PD holds 512 entries */
242 		count = pd_count(size, shift);
243 		while (count--) {
244 			struct i915_page_directory *pd;
245 
246 			pd = alloc_pd(vm);
247 			if (IS_ERR(pd)) {
248 				i915_vm_free_pt_stash(vm, stash);
249 				return PTR_ERR(pd);
250 			}
251 
252 			pd->pt.stash = stash->pt[1];
253 			stash->pt[1] = &pd->pt;
254 		}
255 	}
256 
257 	return 0;
258 }
259 
i915_vm_map_pt_stash(struct i915_address_space * vm,struct i915_vm_pt_stash * stash)260 int i915_vm_map_pt_stash(struct i915_address_space *vm,
261 			 struct i915_vm_pt_stash *stash)
262 {
263 	struct i915_page_table *pt;
264 	int n, err;
265 
266 	for (n = 0; n < ARRAY_SIZE(stash->pt); n++) {
267 		for (pt = stash->pt[n]; pt; pt = pt->stash) {
268 			err = map_pt_dma_locked(vm, pt->base);
269 			if (err)
270 				return err;
271 		}
272 	}
273 
274 	return 0;
275 }
276 
i915_vm_free_pt_stash(struct i915_address_space * vm,struct i915_vm_pt_stash * stash)277 void i915_vm_free_pt_stash(struct i915_address_space *vm,
278 			   struct i915_vm_pt_stash *stash)
279 {
280 	struct i915_page_table *pt;
281 	int n;
282 
283 	for (n = 0; n < ARRAY_SIZE(stash->pt); n++) {
284 		while ((pt = stash->pt[n])) {
285 			stash->pt[n] = pt->stash;
286 			free_px(vm, pt, n);
287 		}
288 	}
289 }
290 
ppgtt_set_pages(struct i915_vma * vma)291 int ppgtt_set_pages(struct i915_vma *vma)
292 {
293 	GEM_BUG_ON(vma->pages);
294 
295 	vma->pages = vma->obj->mm.pages;
296 	vma->page_sizes = vma->obj->mm.page_sizes;
297 
298 	return 0;
299 }
300 
ppgtt_init(struct i915_ppgtt * ppgtt,struct intel_gt * gt)301 void ppgtt_init(struct i915_ppgtt *ppgtt, struct intel_gt *gt)
302 {
303 	struct drm_i915_private *i915 = gt->i915;
304 
305 	ppgtt->vm.gt = gt;
306 	ppgtt->vm.i915 = i915;
307 	ppgtt->vm.dma = i915->drm.dev;
308 	ppgtt->vm.total = BIT_ULL(INTEL_INFO(i915)->ppgtt_size);
309 
310 	dma_resv_init(&ppgtt->vm._resv);
311 	i915_address_space_init(&ppgtt->vm, VM_CLASS_PPGTT);
312 
313 	ppgtt->vm.vma_ops.bind_vma    = ppgtt_bind_vma;
314 	ppgtt->vm.vma_ops.unbind_vma  = ppgtt_unbind_vma;
315 	ppgtt->vm.vma_ops.set_pages   = ppgtt_set_pages;
316 	ppgtt->vm.vma_ops.clear_pages = clear_pages;
317 }
318