xref: /openbmc/linux/drivers/gpu/drm/i915/gt/intel_ppgtt.c (revision 11a163f2)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2020 Intel Corporation
4  */
5 
6 #include <linux/slab.h>
7 
8 #include "i915_trace.h"
9 #include "intel_gtt.h"
10 #include "gen6_ppgtt.h"
11 #include "gen8_ppgtt.h"
12 
13 struct i915_page_table *alloc_pt(struct i915_address_space *vm)
14 {
15 	struct i915_page_table *pt;
16 
17 	pt = kmalloc(sizeof(*pt), I915_GFP_ALLOW_FAIL);
18 	if (unlikely(!pt))
19 		return ERR_PTR(-ENOMEM);
20 
21 	pt->base = vm->alloc_pt_dma(vm, I915_GTT_PAGE_SIZE_4K);
22 	if (IS_ERR(pt->base)) {
23 		kfree(pt);
24 		return ERR_PTR(-ENOMEM);
25 	}
26 
27 	atomic_set(&pt->used, 0);
28 	return pt;
29 }
30 
31 struct i915_page_directory *__alloc_pd(int count)
32 {
33 	struct i915_page_directory *pd;
34 
35 	pd = kzalloc(sizeof(*pd), I915_GFP_ALLOW_FAIL);
36 	if (unlikely(!pd))
37 		return NULL;
38 
39 	pd->entry = kcalloc(count, sizeof(*pd->entry), I915_GFP_ALLOW_FAIL);
40 	if (unlikely(!pd->entry)) {
41 		kfree(pd);
42 		return NULL;
43 	}
44 
45 	spin_lock_init(&pd->lock);
46 	return pd;
47 }
48 
49 struct i915_page_directory *alloc_pd(struct i915_address_space *vm)
50 {
51 	struct i915_page_directory *pd;
52 
53 	pd = __alloc_pd(I915_PDES);
54 	if (unlikely(!pd))
55 		return ERR_PTR(-ENOMEM);
56 
57 	pd->pt.base = vm->alloc_pt_dma(vm, I915_GTT_PAGE_SIZE_4K);
58 	if (IS_ERR(pd->pt.base)) {
59 		kfree(pd->entry);
60 		kfree(pd);
61 		return ERR_PTR(-ENOMEM);
62 	}
63 
64 	return pd;
65 }
66 
67 void free_px(struct i915_address_space *vm, struct i915_page_table *pt, int lvl)
68 {
69 	BUILD_BUG_ON(offsetof(struct i915_page_directory, pt));
70 
71 	if (lvl) {
72 		struct i915_page_directory *pd =
73 			container_of(pt, typeof(*pd), pt);
74 		kfree(pd->entry);
75 	}
76 
77 	if (pt->base)
78 		i915_gem_object_put(pt->base);
79 
80 	kfree(pt);
81 }
82 
83 static inline void
84 write_dma_entry(struct drm_i915_gem_object * const pdma,
85 		const unsigned short idx,
86 		const u64 encoded_entry)
87 {
88 	u64 * const vaddr = kmap_atomic(__px_page(pdma));
89 
90 	vaddr[idx] = encoded_entry;
91 	clflush_cache_range(&vaddr[idx], sizeof(u64));
92 	kunmap_atomic(vaddr);
93 }
94 
95 void
96 __set_pd_entry(struct i915_page_directory * const pd,
97 	       const unsigned short idx,
98 	       struct i915_page_table * const to,
99 	       u64 (*encode)(const dma_addr_t, const enum i915_cache_level))
100 {
101 	/* Each thread pre-pins the pd, and we may have a thread per pde. */
102 	GEM_BUG_ON(atomic_read(px_used(pd)) > NALLOC * I915_PDES);
103 
104 	atomic_inc(px_used(pd));
105 	pd->entry[idx] = to;
106 	write_dma_entry(px_base(pd), idx, encode(px_dma(to), I915_CACHE_LLC));
107 }
108 
109 void
110 clear_pd_entry(struct i915_page_directory * const pd,
111 	       const unsigned short idx,
112 	       const struct drm_i915_gem_object * const scratch)
113 {
114 	GEM_BUG_ON(atomic_read(px_used(pd)) == 0);
115 
116 	write_dma_entry(px_base(pd), idx, scratch->encode);
117 	pd->entry[idx] = NULL;
118 	atomic_dec(px_used(pd));
119 }
120 
121 bool
122 release_pd_entry(struct i915_page_directory * const pd,
123 		 const unsigned short idx,
124 		 struct i915_page_table * const pt,
125 		 const struct drm_i915_gem_object * const scratch)
126 {
127 	bool free = false;
128 
129 	if (atomic_add_unless(&pt->used, -1, 1))
130 		return false;
131 
132 	spin_lock(&pd->lock);
133 	if (atomic_dec_and_test(&pt->used)) {
134 		clear_pd_entry(pd, idx, scratch);
135 		free = true;
136 	}
137 	spin_unlock(&pd->lock);
138 
139 	return free;
140 }
141 
142 int i915_ppgtt_init_hw(struct intel_gt *gt)
143 {
144 	struct drm_i915_private *i915 = gt->i915;
145 
146 	gtt_write_workarounds(gt);
147 
148 	if (IS_GEN(i915, 6))
149 		gen6_ppgtt_enable(gt);
150 	else if (IS_GEN(i915, 7))
151 		gen7_ppgtt_enable(gt);
152 
153 	return 0;
154 }
155 
156 static struct i915_ppgtt *
157 __ppgtt_create(struct intel_gt *gt)
158 {
159 	if (INTEL_GEN(gt->i915) < 8)
160 		return gen6_ppgtt_create(gt);
161 	else
162 		return gen8_ppgtt_create(gt);
163 }
164 
165 struct i915_ppgtt *i915_ppgtt_create(struct intel_gt *gt)
166 {
167 	struct i915_ppgtt *ppgtt;
168 
169 	ppgtt = __ppgtt_create(gt);
170 	if (IS_ERR(ppgtt))
171 		return ppgtt;
172 
173 	trace_i915_ppgtt_create(&ppgtt->vm);
174 
175 	return ppgtt;
176 }
177 
178 void ppgtt_bind_vma(struct i915_address_space *vm,
179 		    struct i915_vm_pt_stash *stash,
180 		    struct i915_vma *vma,
181 		    enum i915_cache_level cache_level,
182 		    u32 flags)
183 {
184 	u32 pte_flags;
185 
186 	if (!test_bit(I915_VMA_ALLOC_BIT, __i915_vma_flags(vma))) {
187 		vm->allocate_va_range(vm, stash, vma->node.start, vma->size);
188 		set_bit(I915_VMA_ALLOC_BIT, __i915_vma_flags(vma));
189 	}
190 
191 	/* Applicable to VLV, and gen8+ */
192 	pte_flags = 0;
193 	if (i915_gem_object_is_readonly(vma->obj))
194 		pte_flags |= PTE_READ_ONLY;
195 
196 	vm->insert_entries(vm, vma, cache_level, pte_flags);
197 	wmb();
198 }
199 
200 void ppgtt_unbind_vma(struct i915_address_space *vm, struct i915_vma *vma)
201 {
202 	if (test_and_clear_bit(I915_VMA_ALLOC_BIT, __i915_vma_flags(vma)))
203 		vm->clear_range(vm, vma->node.start, vma->size);
204 }
205 
206 static unsigned long pd_count(u64 size, int shift)
207 {
208 	/* Beware later misalignment */
209 	return (size + 2 * (BIT_ULL(shift) - 1)) >> shift;
210 }
211 
212 int i915_vm_alloc_pt_stash(struct i915_address_space *vm,
213 			   struct i915_vm_pt_stash *stash,
214 			   u64 size)
215 {
216 	unsigned long count;
217 	int shift, n;
218 
219 	shift = vm->pd_shift;
220 	if (!shift)
221 		return 0;
222 
223 	count = pd_count(size, shift);
224 	while (count--) {
225 		struct i915_page_table *pt;
226 
227 		pt = alloc_pt(vm);
228 		if (IS_ERR(pt)) {
229 			i915_vm_free_pt_stash(vm, stash);
230 			return PTR_ERR(pt);
231 		}
232 
233 		pt->stash = stash->pt[0];
234 		stash->pt[0] = pt;
235 	}
236 
237 	for (n = 1; n < vm->top; n++) {
238 		shift += ilog2(I915_PDES); /* Each PD holds 512 entries */
239 		count = pd_count(size, shift);
240 		while (count--) {
241 			struct i915_page_directory *pd;
242 
243 			pd = alloc_pd(vm);
244 			if (IS_ERR(pd)) {
245 				i915_vm_free_pt_stash(vm, stash);
246 				return PTR_ERR(pd);
247 			}
248 
249 			pd->pt.stash = stash->pt[1];
250 			stash->pt[1] = &pd->pt;
251 		}
252 	}
253 
254 	return 0;
255 }
256 
257 int i915_vm_pin_pt_stash(struct i915_address_space *vm,
258 			 struct i915_vm_pt_stash *stash)
259 {
260 	struct i915_page_table *pt;
261 	int n, err;
262 
263 	for (n = 0; n < ARRAY_SIZE(stash->pt); n++) {
264 		for (pt = stash->pt[n]; pt; pt = pt->stash) {
265 			err = pin_pt_dma(vm, pt->base);
266 			if (err)
267 				return err;
268 		}
269 	}
270 
271 	return 0;
272 }
273 
274 void i915_vm_free_pt_stash(struct i915_address_space *vm,
275 			   struct i915_vm_pt_stash *stash)
276 {
277 	struct i915_page_table *pt;
278 	int n;
279 
280 	for (n = 0; n < ARRAY_SIZE(stash->pt); n++) {
281 		while ((pt = stash->pt[n])) {
282 			stash->pt[n] = pt->stash;
283 			free_px(vm, pt, n);
284 		}
285 	}
286 }
287 
288 int ppgtt_set_pages(struct i915_vma *vma)
289 {
290 	GEM_BUG_ON(vma->pages);
291 
292 	vma->pages = vma->obj->mm.pages;
293 	vma->page_sizes = vma->obj->mm.page_sizes;
294 
295 	return 0;
296 }
297 
298 void ppgtt_init(struct i915_ppgtt *ppgtt, struct intel_gt *gt)
299 {
300 	struct drm_i915_private *i915 = gt->i915;
301 
302 	ppgtt->vm.gt = gt;
303 	ppgtt->vm.i915 = i915;
304 	ppgtt->vm.dma = &i915->drm.pdev->dev;
305 	ppgtt->vm.total = BIT_ULL(INTEL_INFO(i915)->ppgtt_size);
306 
307 	i915_address_space_init(&ppgtt->vm, VM_CLASS_PPGTT);
308 
309 	ppgtt->vm.vma_ops.bind_vma    = ppgtt_bind_vma;
310 	ppgtt->vm.vma_ops.unbind_vma  = ppgtt_unbind_vma;
311 	ppgtt->vm.vma_ops.set_pages   = ppgtt_set_pages;
312 	ppgtt->vm.vma_ops.clear_pages = clear_pages;
313 }
314